if exllama auto diable fused attention

.github/workflows/build_wheels_cuda.yml

@@ -5,12 +5,12 @@ on: workflow_dispatch
 jobs:
   build_wheels:
     if: ${{ github.repository_owner == 'PanQiWei' }}
-    name: Build wheels for ${{ matrix.os }} and Python ${{ matrix.python }} and CUDA ${{ matrix.cuda }}
+    name: Build wheels for ${{ matrix.os }}
     runs-on: ${{ matrix.os }}
     strategy:
       matrix:
-        os: [ubuntu-20.04, windows-latest]
-        pyver: ["3.8", "3.9", "3.10", "3.11"]
+        os: [ubuntu-latest, windows-latest]
+        pyver: ["3.8", "3.9", "3.10"]
         cuda: ["11.7", "11.8"]
     defaults:
       run:
@@ -20,18 +20,20 @@ jobs:
 
     steps:
       - uses: actions/checkout@v3
+        with:
+          ref: 'main'
 
       - uses: actions/setup-python@v3
         with:
           python-version: ${{ matrix.pyver }}
 
-      - name: Setup Miniconda
+      - name: Setup Mamba
         uses: conda-incubator/setup-miniconda@v2.2.0
         with:
           activate-environment: "build"
           python-version: ${{ matrix.pyver }}
           mamba-version: "*"
-          use-mamba: false
+          use-mamba: true
           channels: conda-forge,defaults
           channel-priority: true
           add-pip-as-python-dependency: true
@@ -39,19 +41,14 @@ jobs:
 
       - name: Install Dependencies
         run: |
-          conda install cuda-toolkit -c "nvidia/label/cuda-${env:CUDA_VERSION}.0"
-          conda install pytorch "pytorch-cuda=${env:CUDA_VERSION}" -c pytorch -c nvidia
-          python -m pip install --upgrade build setuptools wheel ninja
+          mamba install -y "pytorch[version=2.0.0,build=py*_cuda${env:CUDA_VERSION}*]" "pytorch-cuda=${env:CUDA_VERSION}" 'sentencepiece' 'cuda' 'ninja' -c 'pytorch' -c "nvidia/label/cuda-${env:CUDA_VERSION}.0" -c 'nvidia' -c 'conda-forge' -c 'defaults'
+          python -m pip install --upgrade build setuptools wheel
 
       - name: Build Wheel
         run: |
           $env:CUDA_PATH = $env:CONDA_PREFIX
           $env:CUDA_HOME = $env:CONDA_PREFIX
           if ($IsLinux) {$env:LD_LIBRARY_PATH = $env:CONDA_PREFIX + '/lib:' + $env:LD_LIBRARY_PATH}
-
-          # TODO: remove this
-          if (!$IsLinux) {$env:INCLUDE_EXLLAMA_KERNELS = 0}
-
           $env:TORCH_CUDA_ARCH_LIST = '6.0 6.1 7.0 7.5 8.0 8.6+PTX'
           if ([decimal]$env:CUDA_VERSION -ge 11.8) { $env:TORCH_CUDA_ARCH_LIST = '6.0 6.1 7.0 7.5 8.0 8.6 8.9 9.0+PTX' }
           python setup.py sdist bdist_wheel
@@ -59,11 +56,11 @@ jobs:
       - uses: actions/upload-artifact@v3
         if: runner.os == 'Linux'
         with:
-          name: 'linux-cuda-wheels'
+          name: 'linux-wheels'
           path: ./dist/*.whl
 
       - uses: actions/upload-artifact@v3
         if: runner.os == 'Windows'
         with:
-          name: 'windows-cuda-wheels'
+          name: 'windows-wheels'
           path: ./dist/*.whl

.github/workflows/build_wheels_pypi.yml

@@ -1,74 +0,0 @@
-name: Build AutoGPTQ Wheels for PyPI with CUDA
-
-on: workflow_dispatch
-
-jobs:
-  build_wheels:
-    if: ${{ github.repository_owner == 'PanQiWei' }}
-    name: Build wheels for ${{ matrix.os }} and Python ${{ matrix.python }} and CUDA 11.7
-    runs-on: ${{ matrix.os }}
-    strategy:
-      matrix:
-        os: [ubuntu-20.04, windows-latest]
-        pyver: ["3.8", "3.9", "3.10", "3.11"]
-    defaults:
-      run:
-        shell: pwsh
-    env:
-        CUDA_VERSION: "11.7"
-
-    steps:
-      - uses: actions/checkout@v3
-
-      - uses: actions/setup-python@v3
-        with:
-          python-version: ${{ matrix.pyver }}
-
-      - name: Setup Miniconda
-        uses: conda-incubator/setup-miniconda@v2.2.0
-        with:
-          activate-environment: "build"
-          python-version: ${{ matrix.pyver }}
-          mamba-version: "*"
-          use-mamba: false
-          channels: conda-forge,defaults
-          channel-priority: true
-          add-pip-as-python-dependency: true
-          auto-activate-base: false
-
-      - name: Install Dependencies
-        run: |
-          conda install cuda-toolkit -c "nvidia/label/cuda-${env:CUDA_VERSION}.0"
-          conda install pytorch "pytorch-cuda=${env:CUDA_VERSION}" -c pytorch -c nvidia
-          python -m pip install --upgrade build setuptools wheel ninja
-
-      - name: Build Wheel
-        run: |
-          $env:CUDA_PATH = $env:CONDA_PREFIX
-          $env:CUDA_HOME = $env:CONDA_PREFIX
-          if ($IsLinux) {$env:LD_LIBRARY_PATH = $env:CONDA_PREFIX + '/lib:' + $env:LD_LIBRARY_PATH}
-
-          $env:TORCH_CUDA_ARCH_LIST = '6.0 6.1 7.0 7.5 8.0 8.6+PTX'
-          if ([decimal]$env:CUDA_VERSION -ge 11.8) { $env:TORCH_CUDA_ARCH_LIST = '6.0 6.1 7.0 7.5 8.0 8.6 8.9 9.0+PTX' }
-
-          $env:PYPI_RELEASE = "1"
-
-          echo "CUDA_PATH:"
-          echo $env:CUDA_PATH
-
-          echo "PYPI_RELEASE:"
-          echo $env:PYPI_RELEASE
-
-          python setup.py sdist bdist_wheel
-
-      - uses: actions/upload-artifact@v3
-        if: runner.os == 'Linux'
-        with:
-          name: 'linux-cuda-wheels'
-          path: ./dist/*.whl
-
-      - uses: actions/upload-artifact@v3
-        if: runner.os == 'Windows'
-        with:
-          name: 'windows-cuda-wheels'
-          path: ./dist/*.whl

.github/workflows/build_wheels_rocm.yml

@@ -8,9 +8,9 @@ jobs:
 
     strategy:
       matrix:
-        os: [ubuntu-20.04]
-        python: ["3.8", "3.9", "3.10", "3.11"]
-        rocm: ["5.4.2"]  # , "5.5", "5.6"]
+        os: [ubuntu-latest]
+        python: ["3.8", "3.9", "3.10"]  # what's the point?
+        rocm: ["5.4.2", "5.5", "5.6"]
 
     name: Build wheels for ${{ matrix.os }} and Python ${{ matrix.python }} and RoCm ${{ matrix.rocm }}
     runs-on: ${{ matrix.os }}
@@ -21,6 +21,12 @@ jobs:
 
     steps:
       - uses: actions/checkout@v3
+        with:
+          ref: 'main'
+
+      - uses: actions/setup-python@v3
+        with:
+          python-version: ${{ matrix.python }}
 
       - name: Free disk space
         run: |
@@ -47,28 +53,8 @@ jobs:
           sudo rm -rf /usr/share/swift > /dev/null 2>&1
           df -h
 
-      - uses: actions/setup-python@v3
-        with:
-          python-version: ${{ matrix.python }}
-
-      - name: Setup Miniconda
-        uses: conda-incubator/setup-miniconda@v2.2.0
-        with:
-          activate-environment: "build"
-          python-version: ${{ matrix.python }}
-          mamba-version: "*"
-          use-mamba: false
-          channels: conda-forge,defaults
-          channel-priority: true
-          add-pip-as-python-dependency: true
-          auto-activate-base: false
-
       - name: Set up environment
         run: |
-          echo "Using python:"
-          python --version
-          which python
-
           if [[ "${{ matrix.rocm }}" == "5.4.2" ]]; then
             export ROCM_DL_FILE=amdgpu-install_5.4.50402-1_all.deb
           elif [[ "${{ matrix.rocm }}" == "5.5" ]]; then
@@ -77,27 +63,20 @@ jobs:
             export ROCM_DL_FILE=amdgpu-install_5.6.50600-1_all.deb
           fi
 
-          curl -O https://repo.radeon.com/amdgpu-install/${{ matrix.rocm }}/ubuntu/focal/$ROCM_DL_FILE
+          curl -O https://repo.radeon.com/amdgpu-install/${{ matrix.rocm }}/ubuntu/jammy/$ROCM_DL_FILE
           sudo dpkg -i $ROCM_DL_FILE
           sudo DEBIAN_FRONTEND=noninteractive amdgpu-install --usecase=rocm --no-dkms --no-32 -y
-
       - name: Install dependencies
         run: |
           sudo apt-get update
           sudo apt-get install -y --no-install-recommends rocsparse-dev rocthrust-dev rocblas-dev hipblas-dev hipsparse-dev
 
+          python -m pip install --pre torch --index-url https://download.pytorch.org/whl/nightly/rocm${{ matrix.rocm }}
           python -m pip install --upgrade build setuptools wheel ninja
-          python -m pip install torch --index-url https://download.pytorch.org/whl/rocm${{ matrix.rocm }}
-
       - name: Build wheels
         run: |
-          echo "Using python for build:"
-          python --version
-          which python
-
           ROCM_VERSION=${{ matrix.rocm }} python setup.py sdist bdist_wheel
-
       - uses: actions/upload-artifact@v3
         with:
-          name: 'linux-rocm-wheels'
+          name: 'wheels'
           path: ./dist/*.whl

README.md

@@ -15,17 +15,10 @@
     </p>
 </h4>
 
-
-## The path to v1.0.0
-
-Hi, fellow community members, long time no see! I'm sorry that I haven't been able to update this project more frequently due to personal reasons during this period. The past few weeks have been huge in terms of my career plans. Not long ago, I officially bid farewell to the startup team that I joined for two years after graduation. I'm very grateful to the leaders and colleagues of the team for their trust and guidance, which enabled me to grow rapidly in two years; at the same time, I'm also really grateful to the team for allowing me to use the internal A100 GPU server cluster free of charge since the start of the AutoGPTQ project to complete various experiments and performance evaluations. (Of course, it can no longer be used in the future, so **it will mean a lot to me if there will be new hardware sponsorship!**) In the past two years, I have served as an AI engineer in this team, responsible for the LLM based dialogue system's architecture design and develop. We had successfully launched a product called gemsouls, but unfortunately it has ceased operations. Now, the team is about to launch a new product called [modelize](https://www.beta.modelize.ai/), which is **a LLM-native AI agent platform, where users can use multiple AI agents to build a highly automated team, allowing them to interact with each other in the workflow, collaborate to complete complex projects efficiently.**
-
-Getting back to the topic, I'm very excited to see that in the past few months, research on optimizing the inference performance of LLMs has made tremendous progress. Now we can not only complete the inference of LLMs on high-end GPUs efficiently, but also on CPUs and even edge devices. A series of technological advancements make me eager to make more contributions to the open source community. Therefore, I will first use about four weeks to gradually update AutoGPTQ to the v1.0.0 official version. During this period, there will also be 2~3 minor versions are released to allow users to experience performance optimization and new features timely. In my vision, **by the time v1.0.0 is officially released, AutoGPTQ will be able to serve as an extendable and flexible quantization backend that supports all GPTQ-like methods and automatically quantize LLMs written by Pytorch**. I detailed the development plan in [this issue](https://github.com/PanQiWei/AutoGPTQ/issues/348), feel free to drop in there for discussion and give your suggestions!
+*<center>📣 Long time no see! 👋 Architecture upgrade, performance optimization and more new features will come in July and August, stay tune! 🥂</center>*
 
 ## News or Update
 
-- 2023-08-23 - (News) - 🤗 Transformers, optimum and peft have integrated `auto-gptq`, so now running and training GPTQ models can be more available to everyone! See [this blog](https://huggingface.co/blog/gptq-integration) and it's resources for more details!
-- 2023-08-21 - (News) - Team of Qwen officially released 4bit quantized version of Qwen-7B based on `auto-gptq`, and provided [a detailed benchmark results](https://huggingface.co/Qwen/Qwen-7B-Chat-Int4#%E9%87%8F%E5%8C%96-quantization)
 - 2023-08-06 - (Update) - Support exllama's q4 CUDA kernel to have at least 1.3x speed up for int4 quantized models when doing inference.
 - 2023-08-04 - (Update) - Support RoCm so that AMD GPU users can use auto-gptq with CUDA extensions.
 - 2023-07-26 - (Update) - An elegant [PPL benchmark script](examples/benchmark/perplexity.py) to get results that can be fairly compared with other libraries such as `llama.cpp`.
@@ -59,17 +52,39 @@ For perplexity comparison, you can turn to [here](https://github.com/qwopqwop200
 ## Installation
 
 ### Quick Installation
-You can install the latest stable release of AutoGPTQ from pip with pre-built wheels compatible with PyTorch 2.0.1:
+You can install the latest stable release of AutoGPTQ from pip:
 
-* For CUDA 11.7: `pip install auto-gptq`
-* For CUDA 11.8: `pip install auto-gptq --extra-index-url https://huggingface.github.io/autogptq-index/whl/cu118/`
-* For RoCm 5.4.2: `pip install auto-gptq --extra-index-url https://huggingface.github.io/autogptq-index/whl/rocm542/`
+```shell
+pip install auto-gptq
+```
 
-**Warning:** These wheels are not expected to work on PyTorch nightly. Please install AutoGPTQ from source when using PyTorch nightly.
+Start from v0.2.0, you can download pre-build wheel that satisfied your environment setup from each version's release assets and install it to skip building stage for the fastest installation speed. For example:
+```shell
+# firstly, cd the directory where the wheel saved, then execute command below
+pip install auto_gptq-0.2.0+cu118-cp310-cp310-linux_x86_64.whl # install v0.2.0 auto_gptq pre-build wheel for linux in an environment whose python=3.10 and cuda=11.8
+```
 
-AutoGPTQ can be installed with the Triton dependency with `pip install auto-gptq[triton]` in order to be able to use the Triton backend (currently only supports linux, no 3-bits quantization).
+#### disable cuda extensions
+By default, cuda extensions will be installed when `torch` and `cuda` is already installed in your machine, if you don't want to use them, using:
+```shell
+BUILD_CUDA_EXT=0 pip install auto-gptq
+```
+And to make sure `autogptq_cuda` is not ever in your virtual environment, run:
+```shell
+pip uninstall autogptq_cuda -y
+```
+
+#### to support triton speedup
+To integrate with `triton`, using:
+> warning: currently triton only supports linux; 3-bit quantization is not supported when using triton
+
+```shell
+pip install auto-gptq[triton]
+```
 
 ### Install from source
+<details>
+<summary>click to see details</summary>
 
 Clone the source code:
 ```shell
@@ -77,18 +92,22 @@ git clone https://github.com/PanQiWei/AutoGPTQ.git && cd AutoGPTQ
 ```
 Then, install from source:
 ```shell
-pip install -v .
+pip install .
 ```
-You can set `BUILD_CUDA_EXT=0` to disable pytorch extension building, but this is **strongly discouraged** as AutoGPTQ then falls back on a slow python implementation.
+Like quick installation, you can also set `BUILD_CUDA_EXT=0` to disable pytorch extension building.
+
+Use `.[triton]` if you want to integrate with triton and it's available on your operating system.
 
 To install from source for AMD GPUs supporting RoCm, please specify the `ROCM_VERSION` environment variable. The compilation can be speeded up by specifying the `PYTORCH_ROCM_ARCH` variable ([reference](https://github.com/pytorch/pytorch/blob/7b73b1e8a73a1777ebe8d2cd4487eb13da55b3ba/setup.py#L132)), for example `gfx90a` for MI200 series devices. Example:
 
 ```
-ROCM_VERSION=5.6 pip install -v .
+ROCM_VERSION=5.6 pip install .
 ```
 
 For RoCm systems, the packages `rocsparse-dev`, `hipsparse-dev`, `rocthrust-dev`, `rocblas-dev` and `hipblas-dev` are required to build.
 
+</details>
+
 ## Quick Tour
 
 ### Quantization and Inference

README_zh.md

@@ -15,16 +15,10 @@
     </p>
 </h4>
 
-## 通向 v1.0.0 之路
-
-嗨，社区的伙伴们，好久不见！很抱歉这段时间由于个人原因，我没能以较高的频率来更新这个项目。过去几周对我的职业生涯规划而言意义重大。在不久前，我正式告别了毕业后便加入两年之久的创业团队，非常感谢团队的领导和同事们给予我的信任与指导，让我能够在两年时间里飞速地成长；同时也十分感激团队允许我自 AutoGPTQ 项目创立以来一直无偿使用内部的 A100 GPU 服务器集群以完成各项实验与性能测评。（当然今后是无法继续使用了，因此**若有新的硬件赞助我将感激不尽**！）过去的两年里，我在这个团队中担任算法工程师的角色，负责基于大语言模型的对话系统架构设计与开发，我们曾成功推出一款名为 gemsouls 的产品，但不幸的是它已经停止运营。而现在，这个团队即将推出一款名为 [modelize](https://www.beta.modelize.ai/) 的新产品，**这是一个大模型原生的 AI 智能体平台，用户可以使用多个 AI 智能体搭建一个高度自动化的团队，让它们在工作流中相互合作，高效完成复杂的项目。**
-
-话归正题，我非常兴奋地看到，在过去几个月的时间里，针对大语言模型推理性能优化的研究取得了巨大的进展，如今我们不仅能够在高端显卡上完成大语言模型的推理，甚至在 CPU 和边缘设备上都可以轻松运行大语言模型。一系列的技术进步，让我同样迫不及待地在开源社区上做出更多的贡献，因此，首先，我将用约四周的时间将 AutoGPTQ 迭代至 v1.0.0 正式版本，在此期间，也会有 2~3 个小版本发布以让用户能够及时体验性能优化和新特性。在我的愿景里，**到 v1.0.0 版本正式发布时，AutoGPTQ 将能够作为一个灵活可拓展的、支持所有 GPTQ-like 方法的量化后端，自动地完成各种基于 Pytorch 编写的大语言模型的量化工作**。我在[这里](https://github.com/PanQiWei/AutoGPTQ/issues/348)详细介绍了开发计划，欢迎移步至此进行讨论并给出你们的建议！
+*<center>📣 好久不见！👋 七月和八月将会迎来架构升级，性能优化和新特性，敬请关注！🥂</center>*
 
 ## 新闻或更新
 
-- 2023-08-23 - (新闻) - 🤗 Transformers、optimum 和 peft 完成了对 `auto-gptq` 的集成，现在使用 GPTQ 模型进行推理和训练将变得更容易！阅读 [这篇博客](https://huggingface.co/blog/gptq-integration) 和相关资源以了解更多细节！
-- 2023-08-21 - (新闻) - 通义千问团队发布了基于 `auto-gptq` 的 Qwen-7B 4bit 量化版本模型，并提供了[详尽的测评结果](https://huggingface.co/Qwen/Qwen-7B-Chat-Int4#%E9%87%8F%E5%8C%96-quantization)
 - 2023-08-06 - (更新) - 支持 exllama 的 q4 CUDA 算子使得 int4 量化模型能够获得至少1.3倍的推理速度提升.
 - 2023-08-04 - (更新) - 支持 RoCm 使得 AMD GPU 的用户能够使用 auto-gptq 的 CUDA 拓展.
 - 2023-07-26 - (更新) - 一个优雅的 [PPL 测评脚本](examples/benchmark/perplexity.py)以获得可以与诸如 `llama.cpp` 等代码库进行公平比较的结果。
@@ -58,14 +52,15 @@
 ## 安装
 
 ### 快速安装
-你可以通过 pip 来安装与 PyTorch 2.0.1 相兼容的最新稳定版本的 AutoGPTQ 的预构建轮子文件：
-
-* 对于 CUDA 11.7： `pip install auto-gptq --extra-index-url https://huggingface.github.io/autogptq-index/whl/cu117/`
-* 对于 CUDA 11.8： `pip install auto-gptq --extra-index-url https://huggingface.github.io/autogptq-index/whl/cu118/`
-* 对于 RoCm 5.4.2： `pip install auto-gptq --extra-index-url https://huggingface.github.io/autogptq-index/whl/rocm542/`
-
-**警告：** 预构建的轮子文件不一定在 PyTorch 的 nightly 版本上有效。如果要使用 PyTorch 的 nightly 版本，请从源码安装 AutoGPTQ。
-
+你可以通过 pip 来安装 AutoGPTQ 当前最新的稳定版本：
+```shell
+pip install auto-gptq
+```
+从 0.2.0 版本开始，你可以从每次版本发布的资产文件列表中下载预构建好的符合你系统配置情况的轮子文件，并通过安装这些轮子文件来跳过漫长的构建过程以达到最快的安装速度。如下是一个例子：
+```shell
+# 首先，进入轮子文件存放的目录，然后执行下面的命令
+pip install auto_gptq-0.2.0+cu118-cp310-cp310-linux_x86_64.whl # 在 linux 操作系统的一个 python=3.10 且 cuda=11.8 的环境下安装 0.2.0 版本的 auto_gptq
+```
 #### 取消 cuda 拓展的安装
 默认情况下，在 `torch` 和 `cuda` 已经于你的机器上被安装时，cuda 拓展将被自动安装，如果你不想要这些拓展的话，采用以下安装命令：
 ```shell

auto_gptq/__init__.py

@@ -1,5 +1,4 @@
-__version__ = "0.5.0.dev0"
+__version__ = "0.3.2"
 from .modeling import BaseQuantizeConfig
 from .modeling import AutoGPTQForCausalLM
 from .utils.peft_utils import get_gptq_peft_model
-from .utils.exllama_utils import exllama_set_max_input_length

auto_gptq/modeling/__init__.py

@@ -12,6 +12,3 @@ from .gpt_bigcode import *
 from .codegen import *
 from .baichuan import *
 from .internlm import *
-from .qwen import *
-from .mistral import *
-from .mpt import *

auto_gptq/modeling/_base.py

@@ -13,7 +13,6 @@ import torch.nn as nn
 import transformers
 from accelerate.hooks import remove_hook_from_module
 from safetensors.torch import save_file as safe_save
-from safetensors.torch import load_file as safe_load
 from transformers import AutoConfig, AutoModelForCausalLM, PreTrainedModel
 from transformers.utils.hub import PushToHubMixin, cached_file, create_repo, create_commit, CommitOperationAdd
 from transformers.utils.generic import ContextManagers
@@ -25,9 +24,7 @@ from ..nn_modules.qlinear import GeneralQuantLinear
 from ..nn_modules._fused_base import FusedBaseAttentionModule, FusedBaseMLPModule
 from ..quantization import GPTQ
 from ..utils.data_utils import collate_data
-from ..utils.import_utils import (
-    dynamically_import_QuantLinear, TRITON_AVAILABLE, AUTOGPTQ_CUDA_AVAILABLE, EXLLAMA_KERNELS_AVAILABLE, QIGEN_AVAILABLE, EXLLAMAV2_KERNELS_AVAILABLE
-)
+from ..utils.import_utils import dynamically_import_QuantLinear, TRITON_AVAILABLE, AUTOGPTQ_CUDA_AVAILABLE
 
 logger = getLogger(__name__)
 
@@ -38,7 +35,7 @@ class BaseQuantizeConfig(PushToHubMixin):
     group_size: int = field(default=-1)
     damp_percent: float = field(default=0.01)
     desc_act: bool = field(default=True)
-    static_groups: bool = field(default=False)
+    static_groups: bool = field(default=True)
     sym: bool = field(default=True)
     true_sequential: bool = field(default=True)
     model_name_or_path: Optional[str] = field(default=None)
@@ -91,17 +88,9 @@ class BaseQuantizeConfig(PushToHubMixin):
                 _commit_hash=commit_hash,
             )
         
-        field_names = [field.name for field in fields(cls)]
         with open(resolved_config_file, "r", encoding="utf-8") as f:
-            args_from_json = json.load(f)
-            filtered_args = {}
-            for key, val in args_from_json.items():
-                if key in field_names:
-                    filtered_args[key] = val
-                else:
-                    logger.warning(f"ignoring unknown parameter in {quantize_config_filename}: {key}.")
-            return cls(**filtered_args)
-
+            return cls(**json.load(f))
+                
     def to_dict(self):
         return {
             "bits": self.bits,
@@ -441,8 +430,7 @@ class BaseGPTQForCausalLM(nn.Module, PushToHubMixin):
             return torch.device(device)
 
     def to(self, device: Union[str, torch.device]):
-        self.model.to(device)
-        return self
+        return self.model.to(device)
 
     def forward(self, *args, **kwargs):
         return self.model(*args, **kwargs)
@@ -697,8 +685,7 @@ class BaseGPTQForCausalLM(nn.Module, PushToHubMixin):
         device: Optional[Union[str, int]] = None,
         low_cpu_mem_usage: bool = False,
         use_triton: bool = False,
-        use_qigen: bool = False,
-        torch_dtype: Optional[torch.dtype] = None,
+        torch_dtype: torch.dtype = torch.float16,
         inject_fused_attention: bool = True,
         inject_fused_mlp: bool = True,
         use_cuda_fp16: bool = True,
@@ -708,8 +695,7 @@ class BaseGPTQForCausalLM(nn.Module, PushToHubMixin):
         trust_remote_code: bool = False,
         warmup_triton: bool = False,
         trainable: bool = False,
-        disable_exllama: bool = True,
-        disable_exllamav2: bool = False,
+        disable_exllama: bool = False,
         **kwargs
     ):
         """load quantized model from local disk"""
@@ -737,53 +723,11 @@ class BaseGPTQForCausalLM(nn.Module, PushToHubMixin):
             "_raise_exceptions_for_missing_entries": False,
             "_commit_hash": commit_hash,
         }
-        if use_qigen and not QIGEN_AVAILABLE:
-            logger.warning("Qigen is not installed, reset use_qigen to False.")
-            use_qigen = False    
+            
         if use_triton and not TRITON_AVAILABLE:
-            logger.warning("Triton is not installed, reset use_triton to False.")
+            logger.warning("triton is not installed, reset use_triton to False")
             use_triton = False
-        if not disable_exllama and not EXLLAMA_KERNELS_AVAILABLE:
-            logger.warning(
-                "Exllama kernel is not installed, reset disable_exllama to True. "
-                "This may because you installed auto_gptq using a pre-build wheel "
-                "on Windows, in which exllama_kernels are not compiled. To use "
-                "exllama_kernels to further speedup inference, you can re-install "
-                "auto_gptq from source."
-            )
-            disable_exllama = True
-        if not disable_exllamav2 and not EXLLAMAV2_KERNELS_AVAILABLE:
-            logger.warning(
-                "Exllamav2 kernel is not installed, reset disable_exllamav2 to True. "
-                "This may because you installed auto_gptq using a pre-build wheel "
-                "on Windows, in which exllama_kernels are not compiled. To use "
-                "exllama_kernels to further speedup inference, you can re-install "
-                "auto_gptq from source."
-            )
-            disable_exllamav2 = True
-        if not AUTOGPTQ_CUDA_AVAILABLE:
-            logger.warning(
-                "CUDA kernels for auto_gptq are not installed, this will result in "
-                "very slow inference speed. This may because:\n"
-                "1. You disabled CUDA extensions compilation by setting BUILD_CUDA_EXT=0 when install auto_gptq from source.\n"
-                "2. You are using pytorch without CUDA support.\n"
-                "3. CUDA and nvcc are not installed in your device."
-            )
-            
-        if use_qigen and QIGEN_AVAILABLE:
-            logger.warning("QIgen is active. Ignores all settings related to cuda.")
-            inject_fused_attention = False
-            inject_fused_mlp = False
-            use_triton = False
-            disable_exllama = True
-            disable_exllamav2 = True
-            
-        if not disable_exllamav2 and not disable_exllama:
-            logger.warning(
-                "You have activated both exllama and exllamav2 kernel. Setting disable_exllama to True and keeping disable_exllamav2 to False"
-            )
-            disable_exllama = True
-            
+
         # == step1: prepare configs and file names == #
         config = AutoConfig.from_pretrained(model_name_or_path, trust_remote_code=trust_remote_code, **cached_file_kwargs)
 
@@ -818,7 +762,7 @@ class BaseGPTQForCausalLM(nn.Module, PushToHubMixin):
                 if isfile(model_save_name + ext):
                     resolved_archive_file = model_save_name + ext
                     break
-        else:  # remote
+        else: # remote
             for ext in extensions:
                 resolved_archive_file = cached_file(model_name_or_path, model_basename + ext, **cached_file_kwargs)
                 if resolved_archive_file is not None:
@@ -829,10 +773,9 @@ class BaseGPTQForCausalLM(nn.Module, PushToHubMixin):
                 
         model_save_name = resolved_archive_file
 
-        if (not disable_exllama or not disable_exllamav2) and trainable:
+        if not disable_exllama and trainable:
             logger.warning("QuantLinear with exllama backend not support trainable mode yet, Switch to the pytorch backend.")
             disable_exllama = True
-            disable_exllamav2 = True
             
         elif not use_triton and trainable:
             logger.warning("QuantLinear with cuda backend not support trainable mode yet, Switch to the pytorch backend.")
@@ -840,95 +783,18 @@ class BaseGPTQForCausalLM(nn.Module, PushToHubMixin):
         # == step2: convert model to gptq-model (replace Linear with QuantLinear) == #
         def skip(*args, **kwargs):
             pass
-            
-        if torch_dtype is None:
-            if not use_qigen:
-                torch_dtype = torch.float16
-            else:
-                torch_dtype = torch.float32
-            
-        if not use_qigen:
-            torch.nn.init.kaiming_uniform_ = skip
-            torch.nn.init.uniform_ = skip
-            torch.nn.init.normal_ = skip
 
-            transformers.modeling_utils._init_weights = False
+        torch.nn.init.kaiming_uniform_ = skip
+        torch.nn.init.uniform_ = skip
+        torch.nn.init.normal_ = skip
 
-            init_contexts = [no_init_weights()]
-            if low_cpu_mem_usage:
-                init_contexts.append(accelerate.init_empty_weights(include_buffers=False))
+        transformers.modeling_utils._init_weights = False
 
-            with ContextManagers(init_contexts):
-                model = AutoModelForCausalLM.from_config(
-                    config,
-                    trust_remote_code=trust_remote_code,
-                    torch_dtype=torch_dtype
-                )
+        init_contexts = [no_init_weights()]
+        if low_cpu_mem_usage:
+            init_contexts.append(accelerate.init_empty_weights(include_buffers=False))
 
-                layers = find_layers(model)
-                ignore_layers = [cls.lm_head_name] + cls.outside_layer_modules
-                for name in list(layers.keys()):
-                    if any([name.startswith(ignore_layer) for ignore_layer in ignore_layers]):
-                        logger.info(f"{name} not been quantized, will be ignored when make_quant.")
-                        del layers[name]
-
-                make_quant(
-                    model,
-                    layers,
-                    quantize_config.bits,
-                    quantize_config.group_size,
-                    use_triton=use_triton,
-                    disable_exllama=disable_exllama,
-                    disable_exllamav2=disable_exllamav2,
-                    use_cuda_fp16=use_cuda_fp16,
-                    desc_act=quantize_config.desc_act,
-                    trainable=trainable
-                )
-                model.tie_weights()
-
-            # == step3: load checkpoint and dispatch == #
-            if isinstance(device_map, str) and device_map not in ["auto", "balanced", "balanced_low_0", "sequential"]:
-                raise ValueError(
-                    "If passing a string for `device_map`, please choose 'auto', 'balanced', 'balanced_low_0' or "
-                    "'sequential'."
-                )
-            if isinstance(device_map, dict):
-                max_memory = None
-            else:
-                if device is None and not device_map and not max_memory:
-                    device_map = "auto"
-                if device is not None:
-                    device = torch.device(device)
-                    if not max_memory and not device_map:
-                        device_map = {"": device.index if device.type == "cuda" else device.type}
-                if not isinstance(device_map, dict) and device_map != "sequential":
-                    max_memory = accelerate.utils.get_balanced_memory(
-                        model=model,
-                        max_memory=max_memory,
-                        no_split_module_classes=[cls.layer_type],
-                        low_zero=(device_map == "balanced_low_0")
-                    )
-            if not isinstance(device_map, dict):
-                device_map = accelerate.infer_auto_device_map(
-                    model,
-                    max_memory=max_memory,
-                    no_split_module_classes=[cls.layer_type]
-                )
-
-            if low_cpu_mem_usage:
-                make_sure_no_tensor_in_meta_device(model, use_triton, quantize_config.desc_act, quantize_config.group_size, bits=quantize_config.bits)
-
-            accelerate.utils.modeling.load_checkpoint_in_model(
-                model,
-                checkpoint=model_save_name,
-                device_map=device_map,
-                offload_state_dict=True,
-                offload_buffers=True
-            )
-            model = simple_dispatch_model(model, device_map)
-        else:
-            if quantize_config.desc_act:
-                NotImplementedError('desc_act=True is not yet supported.')
+        with ContextManagers(init_contexts):
             model = AutoModelForCausalLM.from_config(
                 config,
                 trust_remote_code=trust_remote_code,
@@ -941,11 +807,7 @@ class BaseGPTQForCausalLM(nn.Module, PushToHubMixin):
                 if any([name.startswith(ignore_layer) for ignore_layer in ignore_layers]):
                     logger.info(f"{name} not been quantized, will be ignored when make_quant.")
                     del layers[name]
-            
-            if model_save_name.endswith('.safetensors'):
-                checkpoint = safe_load(model_save_name)
-            else:
-                checkpoint = torch.load(model_save_name)
+
             make_quant(
                 model,
                 layers,
@@ -953,21 +815,54 @@ class BaseGPTQForCausalLM(nn.Module, PushToHubMixin):
                 quantize_config.group_size,
                 use_triton=use_triton,
                 disable_exllama=disable_exllama,
-                disable_exllamav2=disable_exllamav2,
                 use_cuda_fp16=use_cuda_fp16,
                 desc_act=quantize_config.desc_act,
-                trainable=trainable,
-                use_qigen=True
+                trainable=trainable
             )
-            preprocess_checkpoint_qigen(
+            model.tie_weights()
+
+        # == step3: load checkpoint and dispatch == #
+        if isinstance(device_map, str) and device_map not in ["auto", "balanced", "balanced_low_0", "sequential"]:
+            raise ValueError(
+                "If passing a string for `device_map`, please choose 'auto', 'balanced', 'balanced_low_0' or "
+                "'sequential'."
+            )
+        if isinstance(device_map, dict):
+            max_memory = None
+        else:
+            if device is None and not device_map and not max_memory:
+                device_map = "auto"
+            if device is not None:
+                device = torch.device(device)
+                if not max_memory and not device_map:
+                    device_map = {"": device.index if device.type == "cuda" else device.type}
+            if not isinstance(device_map, dict) and device_map != "sequential":
+                max_memory = accelerate.utils.get_balanced_memory(
+                    model=model,
+                    max_memory=max_memory,
+                    no_split_module_classes=[cls.layer_type],
+                    low_zero=(device_map == "balanced_low_0")
+                )
+        if not isinstance(device_map, dict):
+            device_map = accelerate.infer_auto_device_map(
                 model,
-                layers,
-                quantize_config.bits,
-                quantize_config.group_size,
-                checkpoint
+                max_memory=max_memory,
+                no_split_module_classes=[cls.layer_type]
             )
-            model.load_state_dict(checkpoint)
-            # == step4: set seqlen == #
+
+        if low_cpu_mem_usage:
+            make_sure_no_tensor_in_meta_device(model, use_triton, quantize_config.desc_act, quantize_config.group_size, bits=quantize_config.bits)
+
+        accelerate.utils.modeling.load_checkpoint_in_model(
+            model,
+            checkpoint=model_save_name,
+            device_map=device_map,
+            offload_state_dict=True,
+            offload_buffers=True
+        )
+        model = simple_dispatch_model(model, device_map)
+
+        # == step4: set seqlen == #
         model_config = model.config.to_dict()
         seq_len_keys = ["max_position_embeddings", "seq_length", "n_positions"]
         if any([k in model_config for k in seq_len_keys]):
@@ -994,7 +889,6 @@ class BaseGPTQForCausalLM(nn.Module, PushToHubMixin):
                     trainable=trainable,
                     bits=quantize_config.bits,
                     disable_exllama=disable_exllama,
-                    disable_exllamav2=disable_exllamav2
                 )
         if inject_fused_mlp:
             if cls.fused_mlp_module_type is None:
@@ -1069,4 +963,5 @@ class BaseGPTQForCausalLM(nn.Module, PushToHubMixin):
         except:
             return getattr(self.model, item)
 
+
 __all__ = ["BaseGPTQForCausalLM", "BaseQuantizeConfig"]

auto_gptq/modeling/_const.py

@@ -20,17 +20,8 @@ SUPPORTED_MODELS = [
     "RefinedWeb",
     "baichuan",
     "internlm",
-    "qwen",
-    "mpt",
 ]
 if compare_transformers_version("v4.28.0", op="ge"):
     SUPPORTED_MODELS.append("llama")
-if compare_transformers_version("v4.33.0", op="ge"):
-    SUPPORTED_MODELS.append("falcon")
-if compare_transformers_version("v4.34.0", op="ge"):
-    SUPPORTED_MODELS.append("mistral")
 
-
-EXLLAMA_DEFAULT_MAX_INPUT_LENGTH = 2048
-
-__all__ = ["CPU", "CUDA_0", "SUPPORTED_MODELS", "EXLLAMA_DEFAULT_MAX_INPUT_LENGTH"]
+__all__ = ["CPU", "CUDA_0", "SUPPORTED_MODELS"]

auto_gptq/modeling/_utils.py

@@ -1,5 +1,5 @@
 from logging import getLogger
-from typing import Union, Optional
+from typing import Union
 
 import accelerate
 import torch
@@ -7,9 +7,10 @@ import torch.nn as nn
 from transformers import AutoConfig
 import transformers
 
-from ._const import SUPPORTED_MODELS, CPU, CUDA_0, EXLLAMA_DEFAULT_MAX_INPUT_LENGTH
+from ._const import SUPPORTED_MODELS, CPU, CUDA_0
 from ..utils.import_utils import dynamically_import_QuantLinear
 
+
 logger = getLogger(__name__)
 
 
@@ -56,14 +57,12 @@ def make_quant(
     group_size,
     name='',
     use_triton: bool = False,
-    disable_exllama: bool = True,
-    disable_exllamav2: bool = False, 
-    use_qigen: bool = False,
+    disable_exllama: bool = False,
     use_cuda_fp16: bool = True,
     desc_act: bool = False,
     trainable: bool = False
-):  
-    QuantLinear = dynamically_import_QuantLinear(use_triton=use_triton, desc_act=desc_act, group_size=group_size, bits=bits, disable_exllama=disable_exllama, disable_exllamav2=disable_exllamav2, use_qigen=use_qigen)
+):
+    QuantLinear = dynamically_import_QuantLinear(use_triton=use_triton, desc_act=desc_act, group_size=group_size, bits=bits, disable_exllama=disable_exllama)
 
     if isinstance(module, QuantLinear):
         return
@@ -82,7 +81,7 @@ def make_quant(
             elif isinstance(tmp,transformers.pytorch_utils.Conv1D):            
                 in_features = tmp.weight.shape[0]
                 out_features = tmp.weight.shape[1]
-            if (not(desc_act) or group_size == -1) and not use_triton and not use_qigen:
+            if (not(desc_act) or group_size == -1) and not use_triton:
                 new_layer = QuantLinear(
                     bits, group_size, in_features, out_features, True, use_cuda_fp16=use_cuda_fp16, trainable=trainable
                 )
@@ -102,80 +101,8 @@ def make_quant(
             desc_act=desc_act,
             trainable=trainable,
             disable_exllama=disable_exllama,
-            disable_exllamav2=disable_exllamav2,
-            use_qigen=use_qigen
         )
 
-def preprocess_checkpoint_qigen(
-    module,
-    names,
-    bits,
-    group_size,
-    checkpoint,
-    name='',
-):
-    try:
-        import cQIGen as qinfer
-    except ImportError:
-        logger.error('cQIGen not installed.')
-        raise
-
-    QuantLinear = dynamically_import_QuantLinear(use_triton=False, desc_act=False, group_size=group_size, bits=bits, disable_exllama=False, use_qigen=True)
-    if isinstance(module, QuantLinear):
-        in_features = module.infeatures
-        out_features = module.outfeatures
-        
-        zeros = checkpoint[name + '.qzeros']
-        scales = checkpoint[name + '.scales'].float()
-        
-        if zeros.dtype != torch.float32:
-            new_zeros = torch.zeros_like(scales).float().contiguous()
-            if bits == 4:
-                qinfer.unpack_zeros4(zeros, new_zeros, new_zeros.shape[0], new_zeros.shape[1])
-            elif bits == 2:
-                qinfer.unpack_zeros2(zeros, new_zeros, new_zeros.shape[0], new_zeros.shape[1])
-            elif bits == 3:
-                logger.info("Unpacking zeros for 3 bits")
-            new_scales = scales.contiguous()
-        else:
-            if scales.shape[1] != out_features:
-                new_scales = scales.transpose(0,1).contiguous()
-            else:
-                new_scales = scales.contiguous()
-            if zeros.shape[1] != out_features:
-                new_zeros = zeros.transpose(0,1).contiguous()
-            else:
-                new_zeros = zeros.contiguous()
-
-        checkpoint[name + '.zeros'],checkpoint[name + '.scales'] = new_zeros, new_scales
-        del checkpoint[name + '.qzeros']
-        del checkpoint[name + '.g_idx']
-        if name + '.bias' in checkpoint:
-            checkpoint[name + '.bias'] = checkpoint[name + '.bias'].float()
-        else:
-            checkpoint[name + '.bias'] = torch.zeros(out_features)
-        checkpoint_qweight = checkpoint[name + '.qweight'].int().contiguous()
-        if bits == 4:
-            qweight = torch.zeros(int(in_features // 8 * out_features)).int().contiguous()
-            qinfer.pack4(checkpoint_qweight, qweight, in_features // 8, out_features, module.mb, module.tb, module.cutoff)# * (module.tt//tb))
-        elif bits == 3:
-            qweight = torch.zeros(int(in_features // 32 * 3 * out_features)).int().contiguous()
-            qinfer.pack3(checkpoint_qweight, qweight, in_features // 32 * 3, out_features, module.mb // 32 * 3, module.tb, module.cutoff)
-        elif bits == 2:
-            qweight = torch.zeros(int(in_features // 16 * out_features)).int().contiguous()
-            qinfer.pack2(checkpoint_qweight, qweight, in_features // 16, out_features, module.mb, module.tb, module.cutoff)# * (module.tt//tb))
-        checkpoint[name + '.qweight'] = qweight
-        return
-
-    for name1, child in module.named_children():
-        preprocess_checkpoint_qigen(
-            child,
-            names,
-            bits,
-            group_size,
-            checkpoint,
-            name + '.' + name1 if name != '' else name1,
-        )
 
 def pack_model(
     model,
@@ -188,7 +115,7 @@ def pack_model(
     warmup_triton: bool = False,
     force_layer_back_to_cpu: bool = False
 ):
-    QuantLinear = dynamically_import_QuantLinear(use_triton=use_triton, desc_act=desc_act, group_size=group_size, bits=bits, disable_exllama=False, disable_exllamav2=True)
+    QuantLinear = dynamically_import_QuantLinear(use_triton=use_triton, desc_act=desc_act, group_size=group_size, bits=bits)
 
     if force_layer_back_to_cpu:
         model.to(CPU)
@@ -196,7 +123,7 @@ def pack_model(
     logger.info('Packing model...')
     layers = find_layers(model)
     layers = {n: layers[n] for n in quantizers}
-    make_quant(model, quantizers, bits, group_size, use_triton=use_triton, use_cuda_fp16=use_cuda_fp16, desc_act=desc_act, disable_exllama=False, disable_exllamav2=True)
+    make_quant(model, quantizers, bits, group_size, use_triton=use_triton, use_cuda_fp16=use_cuda_fp16, desc_act=desc_act)
     qlayers = find_layers(model, [QuantLinear])
     for name in qlayers:
         logger.info(name)
@@ -260,10 +187,7 @@ def simple_dispatch_model(model, device_map):
     return model
 
 
-def autogptq_post_init(model, use_act_order: bool, max_input_length: Optional[int] = None):
-    """
-    The max_input_length argument is specific to the exllama backend, that requires to initialize a buffer temp_state.
-    """
+def autogptq_post_init(model, use_act_order: bool):
     device_to_buffers_size = {}
 
     model_uses_exllama = False
@@ -278,9 +202,7 @@ def autogptq_post_init(model, use_act_order: bool, max_input_length: Optional[in
                 }
             
             if not use_act_order:
-                submodule._use_act_order = False
-            else:
-                submodule._use_act_order = True
+                submodule.g_idx = None
 
             # Disable this heuristic for detecting act_order, but it could be used instead of the config.
             """
@@ -305,13 +227,9 @@ def autogptq_post_init(model, use_act_order: bool, max_input_length: Optional[in
         device_to_buffers = {}
 
         if use_act_order:
-            if max_input_length is None:
-                max_input_len = EXLLAMA_DEFAULT_MAX_INPUT_LENGTH
-            else:
-                max_input_len = max_input_length
+            # TODO: initialize this properly
+            max_input_len = 2048
         else:
-            if max_input_length is not None:
-                logger.info("Using exllama backend without act-order, the parameter max_input_length was set although not needed, it will be ignored.")
             max_input_len = 1
 
         for device, buffers_size in device_to_buffers_size.items():
@@ -319,9 +237,7 @@ def autogptq_post_init(model, use_act_order: bool, max_input_length: Optional[in
             # The temp_dq buffer is required to dequantize weights when using cuBLAS, typically for the prefill.
             device_to_buffers[device] = {
                 "temp_state": torch.zeros((max_input_len, buffers_size["max_inner_outer_dim"]), dtype=torch.float16, device=device),
-                "temp_dq": torch.zeros((1, buffers_size["max_dq_buffer_size"]), dtype=torch.float16, device=device),
-                "max_dq_buffer_size": buffers_size["max_dq_buffer_size"],
-                "max_inner_outer_dim": buffers_size["max_inner_outer_dim"],
+                "temp_dq": torch.zeros((1, buffers_size["max_dq_buffer_size"]), dtype=torch.float16, device=device)
             }
         
         # Buffers need to be persistent to avoid any bug.
@@ -341,32 +257,8 @@ def autogptq_post_init(model, use_act_order: bool, max_input_length: Optional[in
             if hasattr(submodule, "QUANT_TYPE") and submodule.QUANT_TYPE == "exllama":
                 submodule.post_init()
 
-    ## exllamav2
-    fixed_bytes = {}
-    model_uses_exllamav2 = False
+        torch.cuda.empty_cache()
     
-    for _, submodule in model.named_modules():
-        if hasattr(submodule, "QUANT_TYPE") and submodule.QUANT_TYPE == "exllamav2":
-            model_uses_exllamav2 = True
-            device = submodule.qweight.device
-            scratch_fixed = submodule.scratch_space_fixed()
-            fixed_bytes[device] = max(scratch_fixed, fixed_bytes.get(device,0))
-
-    if model_uses_exllamav2:
-        from ..nn_modules.qlinear.qlinear_exllamav2 import ExLlamaV2DeviceTensors
-        device_tensors = {} 
-        for device, scratch_bytes in fixed_bytes.items():
-            device_tensors[device] = ExLlamaV2DeviceTensors(device.index, scratch_bytes)
-        
-        # have persistent buffers, otherwise we will get OOM
-        model.device_tensors = device_tensors
-
-        for _, submodule in model.named_modules():
-            if hasattr(submodule, "QUANT_TYPE") and submodule.QUANT_TYPE == "exllamav2":
-                device = submodule.qweight.device
-                submodule.post_init(temp_dq = model.device_tensors[device])
-    torch.cuda.empty_cache()
-
     return model
 
 
@@ -384,7 +276,6 @@ __all__ = [
     "get_module_by_name_prefix",
     "get_module_by_name_suffix",
     "make_quant",
-    "preprocess_checkpoint_qigen",
     "pack_model",
     "autogptq_post_init",
     "check_and_get_model_type",

auto_gptq/modeling/auto.py

@@ -15,9 +15,6 @@ from .rw import RWGPTQForCausalLM
 from .gpt_bigcode import GPTBigCodeGPTQForCausalLM
 from .baichuan import BaiChuanGPTQForCausalLM
 from .internlm import InternLMGPTQForCausalLM
-from .qwen import QwenGPTQForCausalLM
-from .mistral import MistralGPTQForCausalLM
-from .mpt import MPTGPTQForCausalLM
 
 GPTQ_CAUSAL_LM_MODEL_MAP = {
     "bloom": BloomGPTQForCausalLM,
@@ -31,12 +28,8 @@ GPTQ_CAUSAL_LM_MODEL_MAP = {
     "codegen": CodeGenGPTQForCausalLM,
     "RefinedWebModel": RWGPTQForCausalLM,
     "RefinedWeb": RWGPTQForCausalLM,
-    "falcon": RWGPTQForCausalLM,
     "baichuan": BaiChuanGPTQForCausalLM,
     "internlm": InternLMGPTQForCausalLM,
-    "qwen": QwenGPTQForCausalLM,
-    "mistral": MistralGPTQForCausalLM,
-    "mpt": MPTGPTQForCausalLM,
 }
 
 
@@ -86,8 +79,6 @@ class AutoGPTQForCausalLM:
         trust_remote_code: bool = False,
         warmup_triton: bool = False,
         trainable: bool = False,
-        disable_exllama: bool = True,
-        disable_exllamav2: bool = False,
         **kwargs
     ) -> BaseGPTQForCausalLM:
         model_type = check_and_get_model_type(model_name_or_path, trust_remote_code)
@@ -127,8 +118,6 @@ class AutoGPTQForCausalLM:
             trust_remote_code=trust_remote_code,
             warmup_triton=warmup_triton,
             trainable=trainable,
-            disable_exllama=disable_exllama,
-            disable_exllamav2=disable_exllamav2,
             **keywords
         )
 

auto_gptq/modeling/mistral.py

@@ -1,16 +0,0 @@
-from ._base import *
-
-
-class MistralGPTQForCausalLM(BaseGPTQForCausalLM):
-    layer_type = "MistralDecoderLayer"
-    layers_block_name = "model.layers"
-    outside_layer_modules = ["model.embed_tokens", "model.norm"]
-    inside_layer_modules = [
-        ["self_attn.k_proj", "self_attn.v_proj", "self_attn.q_proj"],
-        ["self_attn.o_proj"],
-        ["mlp.up_proj", "mlp.gate_proj"],
-        ["mlp.down_proj"],
-    ]
-
-
-__all__ = ["MistralGPTQForCausalLM"]

auto_gptq/modeling/mpt.py

@@ -1,18 +0,0 @@
-from auto_gptq.modeling import BaseGPTQForCausalLM
-
-
-class MPTGPTQForCausalLM(BaseGPTQForCausalLM):
-    layer_type = "MPTBlock"
-    layers_block_name = "transformer.blocks"
-    outside_layer_modules = [
-        "transformer.wte",  "transformer.norm_f"
-    ]
-
-    inside_layer_modules = [
-        ["attn.Wqkv"],
-        ["attn.out_proj"],
-        ["ffn.up_proj"],
-        ["ffn.down_proj"]
-    ]
-
-__all__ = ["MPTGPTQForCausalLM"]

auto_gptq/modeling/qwen.py

@@ -1,16 +0,0 @@
-from ._base import *
-
-
-class QwenGPTQForCausalLM(BaseGPTQForCausalLM):
-    layer_type = "QWenBlock"
-    layers_block_name = "transformer.h"
-    outside_layer_modules = ["transformer.wte", "transformer.wpe", "transformer.ln_f", "transformer.visual"]
-    inside_layer_modules = [
-        ["attn.c_attn"],
-        ["attn.c_proj"],
-        ["mlp.w1", "mlp.w2"],
-        ["mlp.c_proj"]
-    ]
-
-
-__all__ = ["QwenGPTQForCausalLM"]

auto_gptq/nn_modules/fused_gptj_attn.py

@@ -8,6 +8,8 @@ from transformers.models.gptj.modeling_gptj import GPTJAttention
 from ._fused_base import FusedBaseAttentionModule
 from ..utils.import_utils import compare_pytorch_version, dynamically_import_QuantLinear
 
+from logging import getLogger
+logger = getLogger(__name__)
 
 def fixed_pos_embedding(x, seq_dim=1, seq_len=None):
     dim = x.shape[-1]
@@ -235,12 +237,16 @@ class FusedGPTJAttentionForQuantizedModel(FusedBaseAttentionModule):
         desc_act=False,
         trainable=False,
         bits: int = 4,
-        disable_exllama=True,
-        disable_exllamav2=False,
+        disable_exllama=False,
         **kwargs
     ):
         config = model.config
-        QuantLinear = dynamically_import_QuantLinear(use_triton=use_triton, desc_act=desc_act, group_size=group_size, bits=bits, disable_exllama=disable_exllama, disable_exllamav2=disable_exllamav2)
+        QuantLinear = dynamically_import_QuantLinear(use_triton=use_triton, desc_act=desc_act, group_size=group_size, bits=bits, disable_exllama=disable_exllama)
+
+        if QuantLinear.QUANT_TYPE == "exllama" and desc_act:
+            # See fused_llama_attn.py comment
+            logger.warning(f"Exllama kernel does not support query/key/value fusion with act-order. Because of this, Fused attention is automatically disabled.")
+            return False
 
         for name, m in model.named_modules():
             if not isinstance(m, GPTJAttention):
@@ -257,11 +263,7 @@ class FusedGPTJAttentionForQuantizedModel(FusedBaseAttentionModule):
             scales = torch.cat([q_proj.scales, k_proj.scales, v_proj.scales], dim=1)
 
             if QuantLinear.QUANT_TYPE == "exllama":
-                if desc_act:
-                    # See fused_llama_attn.py comment
-                    raise ValueError("Exllama kernel does not support query/key/value fusion with act-order. Please either use inject_fused_attention=False or disable_exllama=True.")
-                else:
-                    g_idx = None
+                g_idx = None
             else:
                 g_idx = torch.cat([q_proj.g_idx, k_proj.g_idx, v_proj.g_idx], dim=0)
 
@@ -299,5 +301,6 @@ class FusedGPTJAttentionForQuantizedModel(FusedBaseAttentionModule):
             setattr(parent, child_name, attn)
             del m
 
+        return True
 
 __all__ = ["FusedGPTJAttentionForQuantizedModel"]

auto_gptq/nn_modules/fused_llama_attn.py

@@ -2,34 +2,48 @@ import math
 import torch
 import torch.nn as nn
 from torch.nn import functional as F
-from transformers.models.llama.modeling_llama import LlamaAttention, apply_rotary_pos_emb
+from transformers.models.llama.modeling_llama import LlamaAttention, apply_rotary_pos_emb, repeat_kv
 
 from ._fused_base import FusedBaseAttentionModule
 from ..utils.import_utils import compare_pytorch_version, dynamically_import_QuantLinear
 
+from logging import getLogger
+logger = getLogger(__name__)
 
 class FusedLlamaAttentionForQuantizedModel(FusedBaseAttentionModule):
     """Multi-headed attention from 'Attention Is All You Need' paper"""
 
     def __init__(
         self,
-        hidden_size,
-        num_heads,
+        config,
         qkv_proj,
         o_proj,
         rotary_emb,
     ):
         super().__init__()
-        self.hidden_size = hidden_size
-        self.num_heads = num_heads
-        self.head_dim = hidden_size // num_heads
+        self.config = config
+        self.hidden_size = config.hidden_size
+        self.num_heads = config.num_attention_heads
+        self.head_dim = self.hidden_size // self.num_heads
+        self.num_key_value_heads = config.num_key_value_heads
+        self.num_key_value_groups = self.num_heads // self.num_key_value_heads
+        self.max_position_embeddings = config.max_position_embeddings
 
-        if self.head_dim * num_heads != self.hidden_size:
+        if (self.head_dim * self.num_heads) != self.hidden_size:
             raise ValueError(
                 f"hidden_size must be divisible by num_heads (got `hidden_size`: {self.hidden_size}"
-                f" and `num_heads`: {num_heads})."
+                f" and `num_heads`: {self.num_heads})."
             )
-        self.qkv_proj = qkv_proj
+        if self.config.pretraining_tp > 1:
+            raise NotImplementedError(f"pretraining_tp of 2 or more is currently not supported.")
+            
+        if len(qkv_proj) == 1:
+            self.qkv_mode = 'qkv'
+            self.qkv_proj = qkv_proj[0]
+        elif len(qkv_proj) == 2:
+            self.qkv_mode = 'q,kv'
+            self.q_proj = qkv_proj[0]
+            self.kv_proj = qkv_proj[1]
         self.o_proj = o_proj
         self.rotary_emb = rotary_emb
 
@@ -39,9 +53,9 @@ class FusedLlamaAttentionForQuantizedModel(FusedBaseAttentionModule):
     def forward(
         self,
         hidden_states,
-        past_key_value=None,
         attention_mask=None,
         position_ids=None,
+        past_key_value=None,
         output_attentions=False,
         use_cache=False,
         **kwargs
@@ -49,14 +63,18 @@ class FusedLlamaAttentionForQuantizedModel(FusedBaseAttentionModule):
         """Input shape: Batch x Time x Channel"""
 
         bsz, q_len, _ = hidden_states.size()
-
-        qkv_states = self.qkv_proj(hidden_states)
-        query_states, key_states, value_states = torch.split(qkv_states, self.hidden_size, dim=2)
-
+        
+        if self.qkv_mode == 'qkv':
+            qkv_states = self.qkv_proj(hidden_states)
+            query_states, key_states, value_states = torch.split(qkv_states, self.hidden_size, dim=2)
+        elif self.qkv_mode == 'q,kv':
+            query_states = self.q_proj(hidden_states)
+            kv_states = self.kv_proj(hidden_states)
+            key_states, value_states = torch.split(kv_states, self.hidden_size, dim=2)
         query_states = query_states.view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
-        key_states = key_states.view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
-        value_states = value_states.view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
-
+        key_states = key_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
+        value_states = value_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
+        
         kv_seq_len = key_states.shape[-2]
         if past_key_value is not None:
             kv_seq_len += past_key_value[0].shape[-2]
@@ -79,7 +97,11 @@ class FusedLlamaAttentionForQuantizedModel(FusedBaseAttentionModule):
 
         past_key_value = (key_states, value_states) if use_cache else None
 
-        if compare_pytorch_version("v2.0.0", op="ge"):
+        # repeat k/v heads if n_kv_heads < n_heads
+        key_states = repeat_kv(key_states, self.num_key_value_groups)
+        value_states = repeat_kv(value_states, self.num_key_value_groups)
+        
+        if compare_pytorch_version("v2.0.0", op="eq"):
             attn_output = F.scaled_dot_product_attention(
                 query_states,
                 key_states,
@@ -93,7 +115,7 @@ class FusedLlamaAttentionForQuantizedModel(FusedBaseAttentionModule):
 
             if attn_weights.size() != (bsz, self.num_heads, q_len, kv_seq_len):
                 raise ValueError(
-                    f"Attention weights should be of size {(bsz * self.num_heads, q_len, kv_seq_len)}, but is"
+                    f"Attention weights should be of size {(bsz, self.num_heads, q_len, kv_seq_len)}, but is"
                     f" {attn_weights.size()}"
                 )
 
@@ -103,7 +125,6 @@ class FusedLlamaAttentionForQuantizedModel(FusedBaseAttentionModule):
                         f"Attention mask should be of size {(bsz, 1, q_len, kv_seq_len)}, but is {attention_mask.size()}"
                     )
                 attn_weights = attn_weights + attention_mask
-                attn_weights = torch.max(attn_weights, torch.tensor(torch.finfo(attn_weights.dtype).min))
 
             # upcast attention to fp32
             attn_weights = nn.functional.softmax(attn_weights, dim=-1, dtype=torch.float32).to(query_states.dtype)
@@ -135,15 +156,20 @@ class FusedLlamaAttentionForQuantizedModel(FusedBaseAttentionModule):
         desc_act=False,
         trainable=False,
         bits: int = 4,
-        disable_exllama=True,
-        disable_exllamav2=False,
+        disable_exllama=False,
         **kwargs
     ):
         """
         Replace all LlamaAttention modules with QuantLlamaAttention modules, fusing the q, k, v projections.
         """
-        QuantLinear = dynamically_import_QuantLinear(use_triton=use_triton, desc_act=desc_act, group_size=group_size, bits=bits, disable_exllama=disable_exllama, disable_exllamav2=disable_exllamav2)
-
+        QuantLinear = dynamically_import_QuantLinear(use_triton=use_triton, desc_act=desc_act, group_size=group_size, bits=bits, disable_exllama=disable_exllama)
+        if QuantLinear.QUANT_TYPE == "exllama" and desc_act:
+            # TODO: support it. The issue lies maybe in the line:
+            # int groups = qzeros.size(0);
+            # in exllama_ext.cpp
+            logger.warning(f"Exllama kernel does not support query/key/value fusion with act-order. Because of this, Fused attention is automatically disabled.")
+            return False
+            
         for name, m in model.named_modules():
             if not isinstance(m, LlamaAttention):
                 continue
@@ -152,41 +178,61 @@ class FusedLlamaAttentionForQuantizedModel(FusedBaseAttentionModule):
             k_proj = m.k_proj
             v_proj = m.v_proj
 
-            qweights = torch.cat([q_proj.qweight, k_proj.qweight, v_proj.qweight], dim=1)
-            qzeros = torch.cat([q_proj.qzeros, k_proj.qzeros, v_proj.qzeros], dim=1)
-            scales = torch.cat([q_proj.scales, k_proj.scales, v_proj.scales], dim=1)
-
-            if QuantLinear.QUANT_TYPE == "exllama":
-                if desc_act:
-                    # TODO: support it. The issue lies maybe in the line:
-                    # int groups = qzeros.size(0);
-                    # in exllama_ext.cpp
-                    raise ValueError("Exllama kernel does not support query/key/value fusion with act-order. Please either use inject_fused_attention=False or disable_exllama=True.")
-                else:
+            if m.num_heads == m.num_key_value_heads:
+                qweights = torch.cat([q_proj.qweight, k_proj.qweight, v_proj.qweight], dim=1)
+                qzeros = torch.cat([q_proj.qzeros, k_proj.qzeros, v_proj.qzeros], dim=1)
+                scales = torch.cat([q_proj.scales, k_proj.scales, v_proj.scales], dim=1)
+                if QuantLinear.QUANT_TYPE == "exllama":
                     g_idx = None
+                else:
+                    g_idx = torch.cat([q_proj.g_idx, k_proj.g_idx, v_proj.g_idx], dim=0)
+                bias = torch.cat([q_proj.bias, k_proj.bias, v_proj.bias], dim=0) if q_proj.bias is not None else None
+
+                qlinear_args = (
+                    q_proj.bits,
+                    q_proj.group_size,
+                    q_proj.infeatures,
+                    q_proj.outfeatures + k_proj.outfeatures + v_proj.outfeatures,
+                    True if q_proj.bias is not None else False,
+                )
+                qlinear_kwargs = {"trainable": trainable}
+                if (not desc_act or group_size == -1) and not use_triton:
+                    qlinear_kwargs["use_cuda_fp16"] = use_cuda_fp16
+                qkv_layer = QuantLinear(*qlinear_args, **qlinear_kwargs)
+                qkv_layer.qweight = qweights
+                qkv_layer.qzeros = qzeros
+                qkv_layer.scales = scales
+                qkv_layer.g_idx = g_idx
+                qkv_layer.bias = bias
+                qkv_layers = [qkv_layer]
             else:
-                g_idx = torch.cat([q_proj.g_idx, k_proj.g_idx, v_proj.g_idx], dim=0)
-            
-            bias = torch.cat([q_proj.bias, k_proj.bias, v_proj.bias], dim=0) if q_proj.bias is not None else None
+                qweights = torch.cat([k_proj.qweight, v_proj.qweight], dim=1)
+                qzeros = torch.cat([k_proj.qzeros, v_proj.qzeros], dim=1)
+                scales = torch.cat([k_proj.scales, v_proj.scales], dim=1)
+                if QuantLinear.QUANT_TYPE == "exllama":
+                    g_idx = None
+                else:
+                    g_idx = torch.cat([k_proj.g_idx, v_proj.g_idx], dim=0)
+                bias = torch.cat([k_proj.bias, v_proj.bias], dim=0) if q_proj.bias is not None else None
 
-            qlinear_args = (
-                q_proj.bits,
-                q_proj.group_size,
-                q_proj.infeatures,
-                q_proj.outfeatures + k_proj.outfeatures + v_proj.outfeatures,
-                True if q_proj.bias is not None else False,
-            )
-            qlinear_kwargs = {"trainable": trainable}
-            if (not desc_act or group_size == -1) and not use_triton:
-                qlinear_kwargs["use_cuda_fp16"] = use_cuda_fp16
-            qkv_layer = QuantLinear(*qlinear_args, **qlinear_kwargs)
-            qkv_layer.qweight = qweights
-            qkv_layer.qzeros = qzeros
-            qkv_layer.scales = scales
-            qkv_layer.g_idx = g_idx
-            qkv_layer.bias = bias
-
-            attn = cls(m.hidden_size, m.num_heads, qkv_layer, m.o_proj, m.rotary_emb)
+                qlinear_args = (
+                    k_proj.bits,
+                    k_proj.group_size,
+                    k_proj.infeatures,
+                    k_proj.outfeatures + v_proj.outfeatures,
+                    True if q_proj.bias is not None else False,
+                )
+                qlinear_kwargs = {"trainable": trainable}
+                if (not desc_act or group_size == -1) and not use_triton:
+                    qlinear_kwargs["use_cuda_fp16"] = use_cuda_fp16
+                kv_layer = QuantLinear(*qlinear_args, **qlinear_kwargs)
+                kv_layer.qweight = qweights
+                kv_layer.qzeros = qzeros
+                kv_layer.scales = scales
+                kv_layer.g_idx = g_idx
+                kv_layer.bias = bias
+                qkv_layers = [q_proj, kv_layer]
+            attn = cls(m.config, qkv_layers, m.o_proj, m.rotary_emb)
 
             if '.' in name:
                 parent_name = name.rsplit('.', 1)[0]
@@ -198,6 +244,5 @@ class FusedLlamaAttentionForQuantizedModel(FusedBaseAttentionModule):
                 child_name = name
 
             setattr(parent, child_name, attn)
-
-
-__all__ = ["FusedLlamaAttentionForQuantizedModel"]
+            
+        return True

auto_gptq/nn_modules/qlinear/qlinear_cuda.py

@@ -219,7 +219,7 @@ class QuantLinear(nn.Module):
                     torch.unsqueeze(self.qzeros, 2).expand(-1, -1, 32 // self.bits),
                     self.wf.unsqueeze(0)
                 ).to(torch.int16 if self.bits == 8 else torch.int8)
-                zeros = torch.bitwise_and(zeros, (2 ** self.bits) - 1)
+                torch.bitwise_and(zeros, (2 ** self.bits) - 1, out=zeros)
 
                 zeros = zeros + 1
                 zeros = zeros.reshape(self.scales.shape)
@@ -228,7 +228,7 @@ class QuantLinear(nn.Module):
                     torch.unsqueeze(self.qweight, 1).expand(-1, 32 // self.bits, -1),
                     self.wf.unsqueeze(-1)
                 ).to(torch.int16 if self.bits == 8 else torch.int8)
-                weight = torch.bitwise_and(weight, (2 ** self.bits) - 1)
+                torch.bitwise_and(weight, (2 ** self.bits) - 1, out=weight)
             elif self.bits == 3:
                 zeros = self.qzeros.reshape(
                     self.qzeros.shape[0], self.qzeros.shape[1] // 3, 3, 1
@@ -267,10 +267,10 @@ class QuantLinear(nn.Module):
                     g_idx_i = self.g_idx[i*num_dim:(i+1)*num_dim]
                     weights.append(scale_i[g_idx_i.long()] * (weight_i - zeros_i[g_idx_i.long()]))
                 weights = torch.cat(weights,dim=1)
-            out = torch.matmul(x.to(weights.dtype), weights)
+            out = torch.matmul(x.half(), weights)
         out = out.half().reshape(out_shape)
         out = out + self.bias if self.bias is not None else out
-        return out.to(x.dtype)
+        return out
 
 
 __all__ = ["QuantLinear"]

auto_gptq/nn_modules/qlinear/qlinear_cuda_old.py

@@ -229,7 +229,7 @@ class QuantLinear(nn.Module):
                 
             if self.bits in [2,4,8]:
                zeros = torch.bitwise_right_shift(torch.unsqueeze(self.qzeros, 2).expand(-1, -1, 32 // self.bits), self.wf.unsqueeze(0)).to(torch.int16 if self.bits == 8 else torch.int8)
-               zeros = torch.bitwise_and(zeros, (2 ** self.bits) - 1)
+               torch.bitwise_and(zeros, (2 ** self.bits) - 1, out=zeros)
                    
                zeros = zeros + 1
                zeros = zeros.reshape(-1, 1, zeros.shape[1] * zeros.shape[2])
@@ -238,7 +238,7 @@ class QuantLinear(nn.Module):
                scales = scales.reshape(-1, 1, scales.shape[-1])
                 
                weight = torch.bitwise_right_shift(torch.unsqueeze(self.qweight, 1).expand(-1, 32 // self.bits, -1), self.wf.unsqueeze(-1)).to(torch.int16 if self.bits == 8 else torch.int8)
-               weight = torch.bitwise_and(weight,(2 ** self.bits) - 1)
+               torch.bitwise_and(weight,(2 ** self.bits) - 1, out=weight)
                weight = weight.reshape(-1, self.group_size, weight.shape[2])
             elif self.bits == 3:
                zeros = self.qzeros.reshape(self.qzeros.shape[0], self.qzeros.shape[1]//3, 3, 1).expand(-1, -1, -1, 12)
@@ -266,10 +266,10 @@ class QuantLinear(nn.Module):
             weight = (scales * (weight - zeros))
             weight = weight.reshape(weight.shape[0] * weight.shape[1], weight.shape[2])
 
-            out = torch.matmul(x.to(weight.dtype), weight)
+            out = torch.matmul(x.half(), weight)
         out = out.half().reshape(out_shape)
         out = out + self.bias if self.bias is not None else out
-        return out.to(x.dtype)
+        return out
 
 
 __all__ = ["QuantLinear"]

auto_gptq/nn_modules/qlinear/qlinear_exllama.py

@@ -1,24 +1,14 @@
 # Adapted from turboderp exllama: https://github.com/turboderp/exllama
 
-from logging import getLogger
-
+from exllama_kernels import make_q4, q4_matmul
 import torch
 import torch.nn as nn
 import math
 import numpy as np
 import transformers
 
-logger = getLogger(__name__)
-
-try:
-    from exllama_kernels import make_q4, q4_matmul
-except ImportError:
-    logger.error('exllama_kernels not installed.')
-    raise
-
 # Dummy tensor to pass instead of g_idx since there is no way to pass "None" to a C++ extension
-none_tensor = torch.empty((1, 1), device="meta")
-
+none_tensor = torch.empty((1, 1), device = "meta")
 
 def ext_make_q4(qweight, qzeros, scales, g_idx, device):
     """Construct Q4Matrix, return handle"""
@@ -28,12 +18,11 @@ def ext_make_q4(qweight, qzeros, scales, g_idx, device):
                    g_idx if g_idx is not None else none_tensor,
                    device)
 
-
 def ext_q4_matmul(x, q4, q4_width):
     """Matrix multiplication, returns x @ q4"""
     outshape = x.shape[:-1] + (q4_width,)
     x = x.view(-1, x.shape[-1])
-    output = torch.empty((x.shape[0], q4_width), dtype=torch.float16, device=x.device)
+    output = torch.empty((x.shape[0], q4_width), dtype = torch.float16, device = x.device)
 
     q4_matmul(x, q4, output)
 
@@ -44,15 +33,21 @@ class QuantLinear(nn.Module):
     QUANT_TYPE = "exllama"
 
     """Linear layer implementation with per-group 4-bit quantization of the weights"""
-
-    def __init__(self, bits, group_size, infeatures, outfeatures, bias, trainable=False, **kwargs):
+    def __init__(self,
+        bits,
+        group_size,
+        infeatures,
+        outfeatures,
+        bias,
+        trainable=False,
+        **kwargs,
+    ):
         super().__init__()
         if bits != 4:
-            raise ValueError(
-                f"Exllama kernel supports only bits=4, requested bits={bits}. Something is wrong in the model initialization.")
+            raise ValueError(f"Exllama kernel supports only bits=4, requested bits={bits}. Something is wrong in the model initialization.")
         if trainable:
             raise NotImplementedError("Exllama kernel does not support training.")
-
+        
         self.infeatures = infeatures
         self.outfeatures = outfeatures
         self.bits = bits
@@ -89,15 +84,15 @@ class QuantLinear(nn.Module):
     def post_init(self):
         assert self.qweight.device.type == "cuda"
         assert self.qweight.device.index is not None
-
+        
         self.width = self.qweight.shape[1]
 
-        # make_q4 segfaults if g_idx is not on cpu in the act-order case. In the non act-order case, None needs to be passed for g_idx.
+        # make_q4 segfaults if g_idx is not on cpu
         self.q4 = ext_make_q4(
             self.qweight,
             self.qzeros,
             self.scales,
-            self.g_idx.to("cpu") if self._use_act_order else None,
+            self.g_idx.to("cpu") if self.g_idx is not None else self.g_idx,
             self.qweight.device.index
         )
 

auto_gptq/nn_modules/qlinear/qlinear_exllamav2.py

@@ -1,188 +0,0 @@
-# Adapted from turboderp exllama: https://github.com/turboderp/exllamav2
-
-from logging import getLogger
-
-import torch
-import torch.nn as nn
-import math
-
-logger = getLogger(__name__)
-
-try:
-    from exllamav2_kernels import make_q_matrix, gemm_half_q_half
-except ImportError:
-    logger.error('exllamav2_kernels not installed.')
-    raise
-
-# Dummy tensor to pass instead of g_idx since there is no way to pass "None" to a C++ extension
-none_tensor = torch.empty((1, 1), device="meta")
-
-def _torch_device(idx):
-    if idx == -1: return "cpu"
-    return f"cuda:{idx}"
-
-def ext_gemm_half_q_half(x, q_handle, q4_width, force_cuda):
-    """Matrix multiplication, returns x @ q4"""
-    output_shape = x.shape[:-1] + (q4_width,)
-    x = x.view(-1, x.shape[-1])
-    output = torch.empty((x.shape[0], q4_width), dtype = torch.half, device = x.device)
-    gemm_half_q_half(x, q_handle, output, force_cuda)
-    return output.view(output_shape)
-
-def ext_make_q_matrix(w: dict, temp_dq, key: str = None):
-    """
-    Create Q matrix 
-    """
-    # EXL2
-    # won't work as the moment because the tensors are not the same. 
-    if "q_weight" in w:
-        w["q_scale_max"] /= 256
-        w["q_perm"] = w["q_perm"].short()
-        w["q_invperm"] = w["q_invperm"].short()
-        return make_q_matrix(w["q_weight"],
-                                w["q_perm"],
-                                w["q_invperm"],
-                                w["q_scale"],
-                                w["q_scale_max"],
-                                w["q_groups"],
-                                none_tensor,
-                                none_tensor,
-                                none_tensor,
-                                temp_dq)
-    # GPTQ
-    elif "qweight" in w:
-        if w["scales"].dtype == torch.float:
-            w["scales"] = w["scales"].half()
-
-        # GPTQ with g_idx (act_order)
-        if "g_idx" in w and not (w["g_idx"] == 0).all().item():
-            w["q_perm"] = torch.empty((w["qweight"].shape[0] * 8,), dtype = torch.short, device = w["qweight"].device)
-            w["q_invperm"] = torch.empty_like(w["q_perm"])
-            # make_q4 segfaults if g_idx is not on cpu in the act-order case. In the non act-order case, None needs to be passed for g_idx.
-            return make_q_matrix(w["qweight"],
-                                 w["q_perm"],
-                                 w["q_invperm"],
-                                 none_tensor,
-                                 none_tensor,
-                                 none_tensor,
-                                 w["qzeros"],
-                                 w["scales"],
-                                 w["g_idx"].cpu(),
-                                 temp_dq)
-        # GPTQ without g_idx
-        else:
-            return make_q_matrix(w["qweight"],
-                                none_tensor,
-                                none_tensor,
-                                none_tensor,
-                                none_tensor,
-                                none_tensor,
-                                w["qzeros"],
-                                w["scales"],
-                                none_tensor,
-                                temp_dq)
-
-class QuantLinear(nn.Module):
-    QUANT_TYPE = "exllamav2"
-
-    """Linear layer implementation with per-group 4-bit quantization of the weights"""
-
-    def __init__(self, bits, group_size, infeatures, outfeatures, bias, trainable=False, **kwargs):
-        super().__init__()
-        if bits != 4:
-            raise ValueError(
-                f"Exllamav2 kernel supports only bits=4, requested bits={bits}. Something is wrong in the model initialization.")
-        if trainable:
-            raise NotImplementedError("Exllamav2 kernel does not support training.")
-
-        self.q_handle = None
-        self.q_tensors = None
-        self.padding = - outfeatures % 32
-        
-        self.infeatures = infeatures
-        self.outfeatures = outfeatures + self.padding
-        self.bits = bits
-        self.group_size = group_size if group_size != -1 else infeatures
-        self.trainable = trainable
-        self.maxq = 2 ** self.bits - 1
-
-        assert infeatures % 32 == 0
-        assert infeatures % self.group_size == 0
-        assert outfeatures % 32 == 0
-        
-        # I need to register the tensors, otherwise, we won't be able to load them easily using transformers ... 
-        self.register_buffer(
-            'qweight',
-            torch.zeros((infeatures // 32 * self.bits, outfeatures), dtype=torch.int32)
-        )
-        self.register_buffer(
-            'qzeros',
-            torch.zeros((math.ceil(infeatures / self.group_size), outfeatures // 32 * self.bits), dtype=torch.int32)
-        )
-        self.register_buffer(
-            'scales',
-            torch.zeros((math.ceil(infeatures / self.group_size), outfeatures), dtype=torch.float16)
-        )
-        self.register_buffer(
-            'g_idx',
-            torch.tensor([i // self.group_size for i in range(infeatures)], dtype=torch.int32)
-        )
-
-        if bias:
-            self.register_buffer('bias', torch.zeros((outfeatures), dtype=torch.float16))
-        else:
-            self.bias = None
-
-    def post_init(self, temp_dq):
-        assert self.qweight.device.type == "cuda"
-        assert self.qweight.device.index is not None
-        self.q_tensors = {
-            "qweight":self.qweight,
-            "qzeros":self.qzeros,
-            "scales":self.scales,
-            "g_idx":self.g_idx
-        }
-        temp_dq = temp_dq.get_scratch_slice(self.temp_dq_size())
-        self.q_handle = ext_make_q_matrix(
-            self.q_tensors, temp_dq
-        )
-    
-    def forward(self, x, force_cuda = False):
-        output = ext_gemm_half_q_half(x, self.q_handle, self.outfeatures, force_cuda)
-
-        if self.bias is not None:
-            output.add_(self.bias)
-        return output
-    
-    def temp_dq_size(self):
-        return self.infeatures * self.outfeatures * 2 + 128
-    
-    def temp_fwd_size(self, max_input_len, max_batch_size):
-        return self.outfeatures * max_input_len * max_batch_size * 4 + 128
-    
-    def scratch_space_fixed(self, max_input_len=2048, max_batch_size=8):
-        return self.temp_dq_size() + self.temp_fwd_size(max_input_len, max_batch_size)
-               
-    
-class ExLlamaV2DeviceTensors:
-
-    device_idx: int
-    scratch_bytes: int
-    scratch_idx: int
-    scratch: torch.tensor = None
-
-    def __init__(self, device_idx, scratch_bytes):
-        self.device_idx = device_idx
-        self.scratch_bytes = scratch_bytes
-    
-    def prepare(self):
-        self.scratch = torch.empty((self.scratch_bytes // 2,), dtype = torch.half, device = _torch_device(self.device_idx))
-
-    def get_scratch_slice(self, size_bytes):
-
-        if self.scratch is None: self.prepare()
-
-        size_bytes = ((size_bytes + 127) // 128) * 128
-        size_half = size_bytes // 2
-        scratch_slice = self.scratch.narrow(0, 0, size_half)
-        return scratch_slice

auto_gptq/nn_modules/qlinear/qlinear_qigen.py

@@ -1,262 +0,0 @@
-from copy import deepcopy
-import torch
-from torch import nn
-from tqdm import tqdm
-import gc
-
-import math
-import numpy as np
-from gekko import GEKKO
-from logging import getLogger
-
-logger = getLogger(__name__)
-
-try:
-    import cQIGen as qinfer
-except ImportError:
-    logger.error('cQIGen not installed.')
-    raise
-
-def mem_model(N, M, T, mu, tu, bits, l1, p, gs):
-    m = GEKKO() # create GEKKO model
-    #cinfergen if bits==3:
-        # tu = tu*3
-    B = m.Const(value=bits)
-    TP = m.Const(value=T//p)
-    k = m.Var(1,integer=True,lb=1)
-    z = m.Var(1,integer=True,lb=1)
-    w = m.Var(1,integer=True,lb=1)
-    y = m.Var(1,integer=True,lb=1)
-    v = m.Var(1,integer=True,lb=1)
-    mb = m.Var(mu,integer=True,lb=1)
-    if gs != -1:
-        gg = m.Var(1,integer=True,lb=1)
-    tb = m.Var(tu,integer=True,lb=1,ub=int(T/p))
-    L = m.Var(integer=True,lb=0,ub=l1)
-    m.Equation(L == 32 * mb * N + B * mb * tb + 32 * tb * N)
-    m.Equation(mb * k == M)
-    if gs != -1:
-        m.Equation(gs * gg == mb)
-    # m.Equation(tb * z == T)
-    m.Equation(tb * z == TP)
-    m.Equation(mu * w == mb)
-    m.Equation(tu * y == tb)
-    # m.Equation(tb * v == tt)
-    m.Maximize(L)
-    m.options.SOLVER = 1
-    m.solver_options = ['minlp_maximum_iterations 1000', \
-                # minlp iterations with integer solution
-                'minlp_max_iter_with_int_sol 10', \
-                # treat minlp as nlp
-                'minlp_as_nlp 0', \
-                # nlp sub-problem max iterations
-                'nlp_maximum_iterations 100', \
-                # 1 = depth first, 2 = breadth first
-                'minlp_branch_method 2', \
-                # maximum deviation from whole number
-                'minlp_integer_tol 0.00', \
-                # covergence tolerance
-                'minlp_gap_tol 0.01']
-    try:
-        m.solve(disp=False)
-    except:
-        try:
-            m.solver_options = ['minlp_maximum_iterations 1000', \
-                            # minlp iterations with integer solution
-                            'minlp_max_iter_with_int_sol 10', \
-                            # treat minlp as nlp
-                            'minlp_as_nlp 0', \
-                            # nlp sub-problem max iterations
-                            'nlp_maximum_iterations 100', \
-                            # 1 = depth first, 2 = breadth first
-                            'minlp_branch_method 1', \
-                            # maximum deviation from whole number
-                            'minlp_integer_tol 0.00', \
-                            # covergence tolerance
-                            'minlp_gap_tol 0.01']
-            m.solve(disp=False)
-        except:
-            # mytb = T//p
-            mytb = tu
-            if gs != -1:
-                mymb = gs
-                while 32 * (mymb + gs) * N + bits * (mymb + gs) * mytb + 32 * mytb * N < l1:
-                    mymb += gs
-                while M % mymb != 0:
-                    mymb -= gs
-                return (int(mymb), int(mytb))
-            else:
-                mymb = mu
-                while 32 * (mymb + mu) * N + bits * (mymb + mu) * mytb + 32 * mytb * N < l1:
-                    mymb += mu
-                while M % mymb != 0:
-                    mymb -= mu
-                return (int(mymb), int(mytb))
-
-    return (int(mb.value[0]), int(tb.value[0]))
-
-params = {}
-
-def compute_reductions(x, gs=-1, cpp=True):
-    if cpp:
-        if len(x.shape) != 1:
-            rows, cols = x.shape
-        else:
-            rows = 1
-            cols = x.shape[0]
-        if gs == -1:
-            out = torch.zeros(rows).float().contiguous()
-            mygs = cols
-        else:
-            out = torch.zeros(rows, cols // gs).float().contiguous()
-            mygs = gs
-        
-        qinfer.compute_reduction_cpp(x, out, rows, cols, mygs)
-        return out
-    if gs == -1: 
-        if len(x.shape) != 1:
-            return torch.sum(x,1)
-        else:
-            return torch.sum(x)
-    else:
-        if len(x.shape) != 1:
-            rows, cols = x.shape
-            out = torch.zeros(rows, cols // gs).float().contiguous()
-            for i in range(cols // gs):
-                out[:,i] = torch.sum(x[:,i*gs:(i+1)*gs],1)
-            return out
-        else:
-            cols = x.shape[0]
-            out = torch.zeros(cols // gs).float().contiguous()
-            for i in range(cols // gs):
-                out[i] = torch.sum(x[i*gs:(i+1)*gs])
-            return out
-
-def process_zeros_scales(zeros, scales, bits, M):
-    if zeros.dtype != torch.float32:
-        new_zeros = torch.zeros_like(scales).float().contiguous()
-        if bits == 4:
-            qinfer.unpack_zeros4(zeros, new_zeros, new_zeros.shape[0], new_zeros.shape[1])
-        elif bits == 2:
-            qinfer.unpack_zeros2(zeros, new_zeros, new_zeros.shape[0], new_zeros.shape[1])
-        elif bits == 3:
-            logger.info("Unpacking zeros for 3 bits")
-        new_scales = scales.contiguous()
-    else:
-        if scales.shape[1] != M:
-            new_scales = scales.transpose(0,1).contiguous()
-        else:
-            new_scales = scales.contiguous()
-        if zeros.shape[1] != M:
-            new_zeros = zeros.transpose(0,1).contiguous()
-        else:
-            new_zeros = zeros.contiguous()
-
-    return new_zeros, new_scales
-    
-class QuantLinear(nn.Module):
-    QUANT_TYPE = "qigen"
-
-    def __init__(self, bits, group_size, infeatures, outfeatures, bias=None, trainable=False, hint=1, p=8, l1=2**18):
-        super().__init__()
-        if bits not in [2, 4]:
-            raise NotImplementedError("Only 2,4 bits are supported.")
-        if trainable:
-            raise NotImplementedError("Qigen kernel does not support training.")
-        self.bits = bits
-        pack = 32 // bits
-
-        self.infeatures = infeatures
-        self.outfeatures = outfeatures
-
-        n = hint
-        m = self.infeatures
-        t = self.outfeatures
-
-        #registers for now are fixed
-        if bits == 3:
-            packed = 32
-            unroll = 3
-            nu = 1 #args.n
-            mu = 32
-            tu = 32
-        else:
-            packed = 32 // bits
-            unroll = 2
-            nu = 1 #args.n
-            mu = 16
-            tu = 32
-        
-        nb = n # it's always small for transformers
-
-        global params
-        if (m,t) in params:
-            mb = params[(m,t)][0]
-            tb = params[(m,t)][1]
-        else:
-            mb, tb = mem_model(n, m, t, mu, tu, bits, l1, p, group_size)
-            params[(m,t)] = (mb,tb)
-
-        split = np.ones(p)
-        split = split * tb
-        while np.sum(split) < t:
-            split = split + tb
-
-        idx = p - 1
-        while np.sum(split) > t:
-            split[idx] = split[idx] - tb
-            idx = idx - 1
-
-        assert(np.sum(split) == t)
-
-        split = split.astype(int)
-        self.tt = int(split[0])
-
-        if split[0] == split[-1]:
-            self.cutoff = int(p+1)
-        else:
-            self.cutoff = int(idx + 1)
-
-        self.mb = mb #// packed
-        self.tb = tb
-
-        self.group_size = group_size
-
-        self.register_buffer('bias', torch.zeros(self.outfeatures))
-        self.register_buffer('zeros', torch.zeros((math.ceil(infeatures / self.group_size), outfeatures), dtype=torch.float32))
-        self.register_buffer('scales', torch.zeros((math.ceil(infeatures / self.group_size), outfeatures), dtype=torch.float32))
-        if bits == 4:
-            self.register_buffer('qweight', torch.zeros(int(self.infeatures // packed * self.outfeatures)).int().contiguous())
-        elif bits == 3:
-            self.register_buffer('qweight', torch.zeros(int(self.infeatures // packed * 3 * self.outfeatures)).int().contiguous())
-        elif bits == 2:
-            self.register_buffer('qweight', torch.zeros(int(self.infeatures // packed * self.outfeatures)).int().contiguous())
-                
-    def forward(self, x):
-        out_shape = x.shape[:-1] + (self.outfeatures,)
-        x = x.reshape((-1, x.shape[-1])).to(torch.float32)
-        B = x.shape[0]
-        new_x = x.T.contiguous()
-        out = torch.zeros((B, self.outfeatures), dtype=torch.float32)
-        sums = compute_reductions(x,gs=self.group_size,cpp=True).contiguous()
-        if self.group_size == -1:
-            if self.bits == 4:
-                qinfer.forward4(new_x, self.qweight, out, self.bias, self.scales, self.zeros, sums, 
-                                B, self.infeatures, self.outfeatures, B, self.mb, self.tb, self.tt, self.cutoff)
-            elif self.bits == 2:
-                qinfer.forward2(new_x, self.qweight, out, self.bias, self.scales, self.zeros, sums, 
-                                B, self.infeatures, self.outfeatures, B, self.mb, self.tb, self.tt, self.cutoff)
-            elif self.bits == 3:
-                qinfer.forward3(new_x, self.qweight, out, self.bias, self.scales, self.zeros, sums, 
-                                B, self.infeatures, self.outfeatures, B, self.mb, self.tb, self.tt, self.cutoff)
-        else:
-            if self.bits == 4:
-                qinfer.forward_gs4(new_x, self.qweight, out, self.bias, self.scales, self.zeros, sums, 
-                                   B, self.infeatures, self.outfeatures, B, self.mb, self.tb, self.tt, self.group_size, self.cutoff)
-            elif self.bits == 2:
-                qinfer.forward_gs2(new_x, self.qweight, out, self.bias, self.scales, self.zeros, sums, 
-                                   B, self.infeatures, self.outfeatures, B, self.mb, self.tb, self.tt, self.group_size, self.cutoff)
-            elif self.bits == 3:
-                qinfer.forward_gs3(new_x, self.qweight, out, self.bias, self.scales, self.zeros, sums,
-                                   B, self.infeatures, self.outfeatures, B, self.mb, self.tb, self.tt, self.group_size, self.cutoff)
-        return out.reshape(out_shape)

auto_gptq/utils/exllama_utils.py

@@ -1,48 +0,0 @@
-import gc
-import torch
-
-def exllama_set_max_input_length(model, max_input_length: int):
-    """
-    This method does not necessarily require `model` to inherit from BaseGPTQForCausalLM.
-
-    When using the exllama backend with act-order, it is necessary to initialize a buffer that depends on the maximum expected input length. In case the
-    default used (EXLLAMA_DEFAULT_MAX_INPUT_LENGTH) is too short, this method can be called to extend the buffer size without reloading the whole model.
-    """
-
-    # The import is set here to avoid a global import. Arguably this is quite ugly, it would be better to have lazy loading.
-    from exllama_kernels import prepare_buffers, cleanup_buffers_cuda
-
-    if not model.quantize_config.desc_act:
-        raise ValueError("The method exllama_set_max_input_length should be called only when using the exllama backend **with act-order**.")
-    
-    device_to_buffers_size = {}
-    for device, buffers in model.device_to_buffers.items():
-        device_to_buffers_size[device] = {"max_dq_buffer_size": buffers["max_dq_buffer_size"], "max_inner_outer_dim": buffers["max_inner_outer_dim"]}
-    
-    # For an unknown reason calling just `del model.device_to_buffers` raises an AttributeError.
-    for key in list(model.device_to_buffers.keys()):
-        del model.device_to_buffers[key]
-    model.device_to_buffers = None
-    del model.device_to_buffers
-
-    gc.collect()
-    torch.cuda.empty_cache()
-    cleanup_buffers_cuda()
-
-    device_to_buffers = {}
-    for device, buffers_size in device_to_buffers_size.items():
-        # The temp_state buffer is required to reorder X in the act-order case.
-        # The temp_dq buffer is required to dequantize weights when using cuBLAS, typically for the prefill.
-        device_to_buffers[device] = {
-            "temp_state": torch.zeros((max_input_length, buffers_size["max_inner_outer_dim"]), dtype=torch.float16, device=device),
-            "temp_dq": torch.zeros((1, buffers_size["max_dq_buffer_size"]), dtype=torch.float16, device=device),
-            "max_dq_buffer_size": buffers_size["max_dq_buffer_size"],
-            "max_inner_outer_dim": buffers_size["max_inner_outer_dim"],
-        }
-
-        prepare_buffers(device, device_to_buffers[device]["temp_state"], device_to_buffers[device]["temp_dq"])
-
-    # Buffers need to be persistent to avoid any bug.
-    model.device_to_buffers = device_to_buffers
-
-    return model

auto_gptq/utils/import_utils.py

@@ -10,56 +10,27 @@ except ImportError:
     TRITON_AVAILABLE = False
 
 try:
-    import autogptq_cuda_256
-    import autogptq_cuda_64
+    import autogptq_cuda
 
     AUTOGPTQ_CUDA_AVAILABLE = True
 except:
     AUTOGPTQ_CUDA_AVAILABLE = False
 
-
-try:
-    import exllama_kernels
-
-    EXLLAMA_KERNELS_AVAILABLE = True
-except:
-    EXLLAMA_KERNELS_AVAILABLE = False
-    
-try:
-    import exllamav2_kernels
-
-    EXLLAMAV2_KERNELS_AVAILABLE = True
-except:
-    EXLLAMAV2_KERNELS_AVAILABLE = False
-    
-try:
-    import cQIGen as qinfer
-
-    QIGEN_AVAILABLE = True
-except:
-    QIGEN_AVAILABLE = False
-
 logger = getLogger(__name__)
 
+def dynamically_import_QuantLinear(use_triton: bool, desc_act: bool, group_size: int, bits: int, disable_exllama: bool = False):
+    if use_triton:
+        if torch.version.hip:
+            logger.warning("Running GPTQ triton version on AMD GPUs is untested and may result in errors or wrong predictions. Please use use_triton=False.")
 
-def dynamically_import_QuantLinear(use_triton: bool, desc_act: bool, group_size: int, bits: int, disable_exllama: bool = True, disable_exllamav2:bool = False, use_qigen: bool = False):
-    if use_qigen:
-        from ..nn_modules.qlinear.qlinear_qigen import QuantLinear
+        from ..nn_modules.qlinear.qlinear_triton import QuantLinear
     else:
-        if use_triton:
-            if torch.version.hip:
-                logger.warning("Running GPTQ triton version on AMD GPUs is untested and may result in errors or wrong predictions. Please use use_triton=False.")
-
-            from ..nn_modules.qlinear.qlinear_triton import QuantLinear
+        if bits == 4 and not disable_exllama:
+            from ..nn_modules.qlinear.qlinear_exllama import QuantLinear
+        elif not desc_act or group_size == -1:
+            from ..nn_modules.qlinear.qlinear_cuda_old import QuantLinear
         else:
-            if bits == 4 and not disable_exllamav2 and EXLLAMAV2_KERNELS_AVAILABLE:
-                from ..nn_modules.qlinear.qlinear_exllamav2 import QuantLinear
-            elif bits == 4 and not disable_exllama and EXLLAMA_KERNELS_AVAILABLE:
-                from ..nn_modules.qlinear.qlinear_exllama import QuantLinear
-            elif not desc_act or group_size == -1:
-                from ..nn_modules.qlinear.qlinear_cuda_old import QuantLinear
-            else:
-                from ..nn_modules.qlinear.qlinear_cuda import QuantLinear
+            from ..nn_modules.qlinear.qlinear_cuda import QuantLinear
 
     return QuantLinear
 

auto_gptq/utils/peft_utils.py

@@ -402,7 +402,7 @@ def get_gptq_peft_model(
     with hijack_peft_mappings():
         try:
             if train_mode:
-                peft_model = get_peft_model(model.model, peft_config, adapter_name=adapter_name)
+                peft_model = get_peft_model(model.model, peft_config)
             else:
                 peft_model = PeftModel.from_pretrained(model.model, model_id, adapter_name)
         except:

autogptq_cuda/exllama/cuda_func/q4_matmul.cu

@@ -240,7 +240,7 @@ void q4_matmul_recons_cuda
     const half* x_mapped = x;
     if (w->cuda_x_map)
     {
-        TORCH_CHECK(buffers->temp_state_size >= x_height * dim, "The temp_state buffer is too small in the exllama backend. Please call the exllama_set_max_input_length function to increase the buffer size. Example:\nfrom auto_gptq import exllama_set_max_input_length\nmodel = exllama_set_max_input_length(model, 4096)");
+        TORCH_CHECK(buffers->temp_state_size >= x_height * dim, "temp_state buffer is too small");
         column_remap_cuda(x, buffers->temp_state, x_height, dim, w->cuda_x_map);
         x_mapped = buffers->temp_state;
     }

autogptq_cuda/exllama/exllama_ext.cpp

@@ -251,5 +251,4 @@ PYBIND11_MODULE(TORCH_EXTENSION_NAME, m)
     m.def("cleanup", &cleanup, "cleanup");
     m.def("make_q4", &make_q4, "make_q4");
     m.def("q4_matmul", &q4_matmul, "q4_matmul");
-    m.def("cleanup_buffers_cuda", &cleanup_buffers_cuda, "cleanup_buffers_cuda");
 }

autogptq_extension/exllamav2/config.h

@@ -1,13 +0,0 @@
-#ifndef _config_h
-#define _config_h
-
-#define MAX_Q_GEMM_ROWS 50
-
-#define QMODE_2BIT 1
-#define QMODE_3BIT 1
-#define QMODE_4BIT 1
-#define QMODE_5BIT 1
-#define QMODE_6BIT 0
-#define QMODE_8BIT 0
-
-#endif

autogptq_extension/exllamav2/cpp/util.h

@@ -1,12 +0,0 @@
-#ifndef _util_h
-#define _util_h
-
-#define DBGS(__x) printf("%s\n", __x)
-#define DBGI(__x) printf("%s: %i\n", #__x, __x)
-#define DBGI2(__x, __y) printf("%s, %s: %i, %i\n", #__x, #__y, __x, __y)
-#define DBGI3(__x, __y, __z) printf("%s, %s, %s: %i, %i, %i\n", #__x, #__y, #__z, __x, __y, __z)
-#define DBGF(__x) printf("%s: %f\n", #__x, __x)
-#define DBGF2(__x, __y) printf("%s, %s: %f, %f\n", #__x, #__y, __x, __y)
-#define DBGF3(__x, __y, __z) printf("%s, %s, %s: %f, %f, %f\n", #__x, #__y, #__z, __x, __y, __z)
-
-#endif

autogptq_extension/exllamav2/cuda/compat.cuh

@@ -1,56 +0,0 @@
-#ifndef _compat_cuh
-#define _compat_cuh
-
-// atomicAdd for half types, to support CC < 7.x
-
-__device__ __forceinline__ void atomicAdd_half(half* address, half val)
-{
-    unsigned int * address_as_ui = (unsigned int *) ((char *)address - ((size_t)address & 2));
-    unsigned int old = *address_as_ui;
-    unsigned int assumed;
-
-    do
-    {
-        assumed = old;
-        __half_raw hsum;
-        hsum.x = (size_t)address & 2 ? (old >> 16) : (old & 0xffff);
-        half tmpres = __hadd(hsum, val);
-        hsum = __half_raw(tmpres);
-        old = (size_t)address & 2 ? (old & 0xffff) | (hsum.x << 16) : (old & 0xffff0000) | hsum.x;
-        old = atomicCAS(address_as_ui, assumed, old);
-    }
-    while (assumed != old);
-}
-
-// atomicAdd for half2 types
-
-__device__ __forceinline__ void atomicAdd_half2(half2* address, half2 val)
-{
-    unsigned int* address_as_ui = (unsigned int*)address;
-    unsigned int old = *address_as_ui;
-    unsigned int assumed;
-    do
-    {
-        assumed = old;
-        half2 old_val = *((half2*)&old);
-        half2 new_val = __hadd2(old_val, val);
-        old = atomicCAS(address_as_ui, assumed, *((unsigned int*)&new_val));
-    }
-    while (assumed != old);
-}
-
-//
-
-#if defined(__CUDA_ARCH__) || defined(USE_ROCM)
-#if __CUDA_ARCH__ < 700 || defined(USE_ROCM)
-
-__device__ __forceinline__ void atomicAdd(half* address, half val) { atomicAdd_half(address, val); }
-
-#if __CUDA_ARCH__ < 600 || defined(USE_ROCM)
-__device__ __forceinline__ void atomicAdd(half2* address, half2 val) { atomicAdd_half2(address, val); }
-#endif
-
-#endif
-#endif
-
-#endif

autogptq_extension/exllamav2/cuda/matrix_view.cuh

@@ -1,121 +0,0 @@
-#ifndef _matrix_view_cuh
-#define _matrix_view_cuh
-
-#include <cuda_runtime.h>
-#include <cuda_fp16.h>
-
-#include "quant/qdq_util.cuh"
-
-class MatrixView_half
-{
-public:
-    const half* data;
-    const int height;
-    const int width;
-
-    __device__ __forceinline__ MatrixView_half(const half* data, const int height, const int width)
-        : data(data), height(height), width(width)
-    { }
-
-    __device__ __forceinline__ half item(int row, int column) const { return data[row * width + column]; }
-    __device__ __forceinline__ half2 item_half2(int row, int column) const { return ((half2*)data)[(row * width + column) / 2]; }
-    __device__ __forceinline__ half2 item_half2half2(int row, int column) const { return __half2half2(data[row * width + column]); }
-    __device__ __forceinline__ const half* item_ptr(int row, int column) const { return &data[row * width + column]; }
-
-    __device__ __forceinline__ void item4(half (&items)[4], int row, int column) const
-    {
-        half2* ptr = (half2*) item_ptr(row, column);
-        half2 i01 = ptr[0];
-        half2 i23 = ptr[1];
-        items[0] = __low2half(i01);
-        items[1] = __high2half(i01);
-        items[2] = __low2half(i23);
-        items[3] = __high2half(i23);
-    }
-    __device__ __forceinline__ void item4_f(float (&items)[4], int row, int column) const
-    {
-        half2* ptr = (half2*)item_ptr(row, column);
-        half2 i01 = ptr[0];
-        half2 i23 = ptr[1];
-        items[0] = __half2float(__low2half(i01));
-        items[1] = __half2float(__high2half(i01));
-        items[2] = __half2float(__low2half(i23));
-        items[3] = __half2float(__high2half(i23));
-    }
-
-    __device__ __forceinline__ void item4_h2(half2 (&items)[4], int row, int column) const
-    {
-        half2* ptr = (half2*)item_ptr(row, column);
-        half2 i01 = ptr[0];
-        half2 i23 = ptr[1];
-        items[0] = __half2half2(__low2half(i01));
-        items[1] = __half2half2(__high2half(i01));
-        items[2] = __half2half2(__low2half(i23));
-        items[3] = __half2half2(__high2half(i23));
-    }
-};
-
-class MatrixView_half_rw
-{
-public:
-    half* data;
-    const int height;
-    const int width;
-
-    __device__ __forceinline__ MatrixView_half_rw(half* data, const int height, const int width)
-        : data(data), height(height), width(width)
-    { }
-
-    __device__ __forceinline__ half item(int row, int column) const { return data[row * width + column]; }
-    __device__ __forceinline__ half2 item_half2(int row, int column) const { return ((half2*)data)[(row * width + column) / 2]; }
-    __device__ __forceinline__ half* item_ptr(int row, int column) { return &data[row * width + column]; }
-    __device__ __forceinline__ void set(int row, int column, half value) { data[row * width + column] = value; }
-    __device__ __forceinline__ void set_half2(int row, int column, half2 value) { ((half2*)data)[(row * width + column) / 2] = value; }
-
-    __device__ __forceinline__ void set4(int row, int column, half v0, half v1, half v2, half v3)
-    {
-        half2 v01 = __halves2half2(v0, v1);
-        half2 v23 = __halves2half2(v2, v3);
-        half2* ptr = (half2*) item_ptr(row, column);
-        ptr[0] = v01;
-        ptr[1] = v23;
-    }
-};
-
-class MatrixView_q4_row
-{
-public:
-    const uint32_t* data;
-    const int height;
-    const int width;
-
-    __device__ __forceinline__ MatrixView_q4_row(const uint32_t* data, const int height, const int width)
-        : data(data), height(height), width(width)
-    { }
-
-    __device__ __forceinline__ int item(int row, int column) const
-    {
-        int shift = (column & 0x07) * 4;
-        return (data[row * width / 8 + column / 8] >> shift) & 0x0f;
-    }
-
-    __device__ __forceinline__ void item2(int (&items)[2], int row, int column) const
-    {
-        int shift = (column & 0x07) * 4;
-        uint32_t d = data[row * width / 8 + column / 8] >> shift;
-        items[0] = d & 0x0f;
-        items[1] = (d >> 4) & 0x0f;
-    }
-
-    __device__ __forceinline__ void item4(int (&items)[4], int row, int column) const
-    {
-        int shift = (column & 0x07) * 4;
-        uint32_t d = data[row * width / 8 + column / 8] >> shift;
-        items[0] = d & 0x0f;
-        items[1] = (d >> 4) & 0x0f;
-        items[2] = (d >> 8) & 0x0f;
-        items[3] = (d >> 12) & 0x0f;
-    }
-};
-
-#endif

autogptq_extension/exllamav2/cuda/q_gemm.cu

@@ -1,238 +0,0 @@
-#include "q_gemm.cuh"
-#include "util.cuh"
-#include "matrix_view.cuh"
-#include "../config.h"
-
-#include "quant/qdq_2.cuh"
-#include "quant/qdq_3.cuh"
-#include "quant/qdq_4.cuh"
-#include "quant/qdq_5.cuh"
-#include "quant/qdq_6.cuh"
-#include "quant/qdq_8.cuh"
-
-#define BLOCK_KN_SIZE 128
-#define BLOCK_M_SIZE_MAX 8
-#define MAX_GROUPS_IN_BLOCK (BLOCK_KN_SIZE / 32)
-#define CLEAR_N_SIZE 256
-
-#include "q_gemm_kernel.cuh"
-#include "q_gemm_kernel_gptq.cuh"
-
-#if defined(USE_ROCM)
-__host__ __forceinline__ hipblasStatus_t __compat_hipblasHgemm(hipblasHandle_t    handle,
-                                                               hipblasOperation_t transA,
-                                                               hipblasOperation_t transB,
-                                                               int                m,
-                                                               int                n,
-                                                               int                k,
-                                                               const half*        alpha,
-                                                               const half*        AP,
-                                                               int                lda,
-                                                               const half*        BP,
-                                                               int                ldb,
-                                                               const half*        beta,
-                                                               half*              CP,
-                                                               int                ldc) {
-    return hipblasHgemm(handle, transA, transB, m, n, k,
-                        reinterpret_cast<const hipblasHalf *>(alpha),
-                        reinterpret_cast<const hipblasHalf *>(AP), lda,
-                        reinterpret_cast<const hipblasHalf *>(BP), ldb,
-                        reinterpret_cast<const hipblasHalf *>(beta),
-                        reinterpret_cast<hipblasHalf *>(CP), ldc);
-}
-#define hipblasHgemm __compat_hipblasHgemm
-
-// Previous version of PyTorch were converting to rocBLAS instead of hipBLAS.
-#define rocblas_operation_none HIPBLAS_OP_N
-#define rocblas_hgemm __compat_hipblasHgemm
-#endif
-
-void gemm_half_q_half_cuda_part
-(
-    const half* a,
-    QMatrix* b,
-    half* c,
-    int size_m,
-    int size_n,
-    int size_k,
-    int m_count,
-    bool clear
-)
-{
-    if (!b->is_gptq)
-    {
-        dim3 blockDim, gridDim;
-        blockDim.x = BLOCK_KN_SIZE;
-        blockDim.y = 1;
-        blockDim.z = 1;
-        gridDim.x = DIVIDE(size_n, BLOCK_KN_SIZE * 4);
-        gridDim.y = DIVIDE(size_m, m_count);
-        gridDim.z = DIVIDE(size_k, BLOCK_KN_SIZE);
-
-        fp_gemm_half_q_half_kernel kernel = pick_gemm_half_q_half_kernel(true, m_count);
-
-        kernel<<<gridDim, blockDim>>>
-        (
-            a,
-            b->cuda_q_weight,
-            b->cuda_q_scale,
-            b->cuda_q_scale_max,
-            c,
-            size_m,
-            size_n,
-            size_k,
-            b->groups,
-            b->groupsize,
-            b->cuda_q_perm,
-            b->rows_8,
-            b->rows_6,
-            b->rows_5,
-            b->rows_4,
-            b->rows_3,
-            b->rows_2,
-            clear
-        );
-    }
-    else
-    {
-        dim3 blockDim, gridDim;
-        blockDim.x = BLOCK_KN_SIZE;
-        blockDim.y = 1;
-        blockDim.z = 1;
-        gridDim.x = DIVIDE(size_n, BLOCK_KN_SIZE * 4);
-        gridDim.y = DIVIDE(size_m, m_count);
-        gridDim.z = DIVIDE(size_k, BLOCK_KN_SIZE);
-
-        fp_gemm_half_q_half_gptq_kernel kernel = pick_gemm_half_q_half_gptq_kernel(true, m_count);
-
-//         DBGX((uint64_t) b->cuda_q_perm);
-//         DBGI(b->rows_4);
-//         DBGI(b->height);
-
-        kernel<<<gridDim, blockDim>>>
-        (
-            a,
-            b->cuda_q_weight,
-            b->cuda_gptq_qzeros,
-            b->cuda_gptq_scales,
-            c,
-            size_m,
-            size_n,
-            size_k,
-            b->groups,
-            b->groupsize,
-            b->cuda_q_perm,
-            b->rows_4,
-            clear
-        );
-    }
-}
-
-void gemm_half_q_half_cuda
-(
-    cublasHandle_t cublas_handle,
-    const half* a,
-    QMatrix* b,
-    half* c,
-    int size_m,
-    int size_n,
-    int size_k,
-    bool clear,
-    half* temp_dq,
-    bool force_cuda
-)
-{
-    if (size_m > MAX_Q_GEMM_ROWS && !force_cuda)
-    {
-        //printf("cublas\n");
-
-        // Reconstruct FP16 matrix, then cuBLAS
-
-        if (!temp_dq) temp_dq = b->temp_dq;
-        b->reconstruct(temp_dq);
-
-        //cublasSetMathMode(cublas_handle, CUBLAS_TENSOR_OP_MATH);
-
-        const half alpha = __float2half(1.0f);
-        const half beta = clear ? __float2half(0.0f) : __float2half(1.0f);
-        cublasHgemm(cublas_handle,
-                    CUBLAS_OP_N,
-                    CUBLAS_OP_N,
-                    size_n, size_m, size_k,
-                    &alpha, temp_dq, size_n,
-                            a,       size_k,
-                    &beta,  c,       size_n);
-
-        //const float alpha = 1.0f;
-        //const float beta = clear ? 0.0f : 1.0f;
-        //cublasSgemmEx(cublas_handle,
-        //              CUBLAS_OP_N,
-        //              CUBLAS_OP_N,
-        //              size_n, size_m, size_k,
-        //              &alpha, temp_dq, CUDA_R_16F, size_n,
-        //                      a,       CUDA_R_16F, size_k,
-        //              &beta,  c,       CUDA_R_16F, size_n);
-
-        //const float alpha = 1.0f;
-        //const float beta = clear ? 0.0f : 1.0f;
-        //cublasGemmEx(cublas_handle,
-        //             CUBLAS_OP_N, CUBLAS_OP_N,
-        //             size_n, size_m, size_k,
-        //             &alpha, temp_dq, CUDA_R_16F, size_n,
-        //                     a,       CUDA_R_16F, size_k,
-        //             &beta,  c,       CUDA_R_16F, size_n,
-        //             CUDA_R_16F, CUBLAS_GEMM_DFALT_TENSOR_OP);
-    }
-    else
-    {
-        //printf("cuda\n");
-
-        // Quantized matmul
-
-        //if (clear) clear_tensor_cuda(c, size_m, size_n);
-
-        int max_chunks = size_m / BLOCK_M_SIZE_MAX;
-        int last_chunk = max_chunks * BLOCK_M_SIZE_MAX;
-        int last_chunk_size = size_m - last_chunk;
-
-        if (max_chunks)
-        {
-            gemm_half_q_half_cuda_part(a, b, c, last_chunk, size_n, size_k, BLOCK_M_SIZE_MAX, clear);
-        }
-
-        if (last_chunk_size)
-        {
-            gemm_half_q_half_cuda_part(a + last_chunk * size_k, b, c + last_chunk * size_n, last_chunk_size, size_n, size_k, last_chunk_size, clear);
-        }
-    }
-}
-
-__global__ void clear_kernel
-(
-    half* __restrict__ c,
-    const int size_m,
-    const int size_n
-)
-{
-    int m = blockIdx.y;
-    int n = (blockIdx.x * CLEAR_N_SIZE + threadIdx.x) * 8;
-    if (n >= size_n) return;
-    int4* c_ptr = (int4*)(c + m * size_n + n);
-    *c_ptr = {};
-}
-
-void clear_tensor_cuda
-(
-    half* c,
-    int size_m,
-    int size_n
-)
-{
-    return;
-    dim3 blockDim, gridDim;
-    blockDim.x = CLEAR_N_SIZE;
-    blockDim.y = 1;
-    gridDim.x = DIVIDE(size_n / 8, CLEAR_N_SIZE);
-    gridDim.y = size_m;
-    clear_kernel<<<gridDim, blockDim>>>(c, size_m, size_n);
-}

autogptq_extension/exllamav2/cuda/q_gemm.cuh

@@ -1,33 +0,0 @@
-#ifndef _q_gemm_cuh
-#define _q_gemm_cuh
-
-#include <cuda_runtime.h>
-#include <cuda_fp16.h>
-#include <cstdint>
-#include <cstdio>
-#include <ATen/cuda/CUDAContext.h>
-
-#include "q_matrix.cuh"
-
-void gemm_half_q_half_cuda
-(
-    cublasHandle_t cublas_handle,
-    const half* a,
-    QMatrix* b,
-    half* c,
-    int size_m,
-    int size_n,
-    int size_k,
-    bool clear = false,
-    half* reconstruct = NULL,
-    bool force_cuda = false
-);
-
-void clear_tensor_cuda
-(
-    half* c,
-    int size_m,
-    int size_n
-);
-
-#endif

autogptq_extension/exllamav2/cuda/q_gemm_kernel.cuh

@@ -1,484 +0,0 @@
-#include "compat.cuh"
-
-__forceinline__ __device__ half2 dot22_8(half2(&dq)[4], const half* a_ptr, const half2 g_result, const half qs_h)
-{
-    half2 result = {};
-    const half2* a2_ptr = (const half2*)a_ptr;
-    #pragma unroll
-    for (int i = 0; i < 4; i++) result = __hfma2(dq[i], *a2_ptr++, result);
-    return __hfma2(result, __halves2half2(qs_h, qs_h), g_result);
-}
-
-__forceinline__ __device__ half2 dot22_16(half2(&dq)[8], const half* a_ptr, const half2 g_result, const half qs_h)
-{
-    half2 result = {};
-    const half2* a2_ptr = (const half2*)a_ptr;
-    #pragma unroll
-    for (int i = 0; i < 8; i++) result = __hfma2(dq[i], *a2_ptr++, result);
-    return __hfma2(result, __halves2half2(qs_h, qs_h), g_result);
-}
-
-__forceinline__ __device__ half2 dot22_32(half2(&dq)[16], const half* a_ptr, const half2 g_result, const half qs_h)
-{
-    half2 result = {};
-    const half2* a2_ptr = (const half2*)a_ptr;
-    #pragma unroll
-    for (int i = 0; i < 16; i += 1) result = __hfma2(dq[i], *a2_ptr++, result);
-    return __hfma2(result, __halves2half2(qs_h, qs_h), g_result);
-}
-
-__forceinline__ __device__ float dot22_8_f(half2(&dq)[4], const half* a_ptr, const float g_result, const float qs_f)
-{
-    half2 result = {};
-    const half2* a2_ptr = (const half2*)a_ptr;
-    #pragma unroll
-    for (int i = 0; i < 4; i++) result = __hfma2(dq[i], *a2_ptr++, result);
-    float result_f = __half2float(__low2half(result)) + __half2float(__high2half(result));
-    return fma(result_f, qs_f, g_result);
-}
-
-__forceinline__ __device__ float dot22_16_f(half2(&dq)[8], const half* a_ptr, const float g_result, const float qs_f)
-{
-    half2 result = {};
-    const half2* a2_ptr = (const half2*)a_ptr;
-    #pragma unroll
-    for (int i = 0; i < 8; i++) result = __hfma2(dq[i], *a2_ptr++, result);
-    float result_f = __half2float(__low2half(result)) + __half2float(__high2half(result));
-    return fma(result_f, qs_f, g_result);
-}
-
-__forceinline__ __device__ float dot22_32_f(half2(&dq)[16], const half* a_ptr, const float g_result, const float qs_f)
-{
-    half2 result = {};
-    const half2* a2_ptr = (const half2*)a_ptr;
-    #pragma unroll
-    for (int i = 0; i < 16; i += 1) result = __hfma2(dq[i], *a2_ptr++, result);
-    float result_f = __half2float(__low2half(result)) + __half2float(__high2half(result));
-    return fma(result_f, qs_f, g_result);
-}
-
-
-
-typedef void (*fp_gemm_half_q_half_kernel)
-(
-    const half*,
-    const uint32_t*,
-    const uint32_t*,
-    const half*,
-    half*,
-    const int,
-    const int,
-    const int,
-    const int,
-    const int,
-    const uint16_t*,
-    const int,
-    const int,
-    const int,
-    const int,
-    const int,
-    const int,
-    const bool
-);
-
-template <bool first_block, int m_count>
-__global__ void gemm_half_q_half_kernel
-(
-    const half*      __restrict__ a,
-    const uint32_t*  __restrict__ b_q_weight,
-    const uint32_t*  __restrict__ b_q_scale,
-    const half*      __restrict__ b_q_scale_max,
-    half*            __restrict__ c,
-    const int size_m,
-    const int size_n,
-    const int size_k,
-    const int groups,
-    const int groupsize,
-    const uint16_t* __restrict__ b_q_perm,
-    const int rows_8,
-    const int rows_6,
-    const int rows_5,
-    const int rows_4,
-    const int rows_3,
-    const int rows_2,
-    const bool clear
-)
-{
-    MatrixView_half a_(a, size_m, size_k);
-    MatrixView_half_rw c_(c, size_m, size_n);
-    MatrixView_q4_row b_q_scale_(b_q_scale, groups, size_n);
-
-    int t = threadIdx.x;
-
-    // Block
-
-    int offset_n = blockIdx.x * BLOCK_KN_SIZE * 4;
-    int offset_m = blockIdx.y * m_count;
-    int offset_k = blockIdx.z * BLOCK_KN_SIZE;
-
-    int end_n = min(offset_n + BLOCK_KN_SIZE * 4, size_n);
-    int end_m = min(offset_m + m_count, size_m);
-    int end_k = min(offset_k + BLOCK_KN_SIZE, size_k);
-    int n = offset_n + t * 4;
-
-    // Preload block_a
-
-    __shared__ half block_a[m_count][BLOCK_KN_SIZE];
-
-    if (offset_k + t < end_k)
-    {
-        for (int m = 0; m < m_count; ++m)
-        {
-            const half* a_ptr = a_.item_ptr(offset_m + m, 0);
-            half* block_a_ptr = block_a[m];
-            half a0 = a_ptr[b_q_perm[offset_k + t]];
-            block_a_ptr[t] = a0;
-        }
-    }
-
-    // Clear
-
-    if (n >= size_n) return;
-
-    if (clear && blockIdx.z == 0) // && (threadIdx.x & 1) == 0)
-    {
-        for (int m = 0; m < m_count; m++)
-            *((uint64_t*) c_.item_ptr(offset_m + m, n)) = 0;
-    }
-
-    __syncthreads();
-
-    // Find initial group
-
-    int group = offset_k / groupsize;
-
-    // Preload scales
-
-    float scales[MAX_GROUPS_IN_BLOCK][4];
-
-    int groups_in_block = DIVIDE((end_k - offset_k), groupsize);
-    for (int g = 0; g < groups_in_block; g++)
-    {
-        int qscales[4];
-        b_q_scale_.item4(qscales, group + g, n);
-        qscales[0]++;
-        qscales[1]++;
-        qscales[2]++;
-        qscales[3]++;
-        float maxscale = __half2float(b_q_scale_max[group + g]);
-        scales[g][0] = __int2float_rn(qscales[0] * qscales[0]) * maxscale;
-        scales[g][1] = __int2float_rn(qscales[1] * qscales[1]) * maxscale;
-        scales[g][2] = __int2float_rn(qscales[2] * qscales[2]) * maxscale;
-        scales[g][3] = __int2float_rn(qscales[3] * qscales[3]) * maxscale;
-    }
-
-    // a, b offset
-
-    int pre_rows_8 = min(rows_8, offset_k);
-    int pre_rows_6 = offset_k > rows_8 ? min(rows_6, offset_k) - rows_8 : 0;
-    int pre_rows_5 = offset_k > rows_6 ? min(rows_5, offset_k) - rows_6 : 0;
-    int pre_rows_4 = offset_k > rows_5 ? min(rows_4, offset_k) - rows_5 : 0;
-    int pre_rows_3 = offset_k > rows_4 ? min(rows_3, offset_k) - rows_4 : 0;
-    int pre_rows_2 = offset_k > rows_3 ? min(rows_2, offset_k) - rows_3 : 0;
-    int qk = 0;
-    qk += pre_rows_8 / 32 * 8;
-    qk += pre_rows_6 / 32 * 6;
-    qk += pre_rows_5 / 32 * 5;
-    qk += pre_rows_4 / 32 * 4;
-    qk += pre_rows_3 / 32 * 3;
-    qk += pre_rows_2 / 32 * 2;
-
-    const uint32_t* b_ptr = b_q_weight + qk * size_n + n;
-    const half* a_ptr = &block_a[0][0];
-    int a_stride = BLOCK_KN_SIZE;
-
-    // Initial group
-
-    int scales_idx = 0;
-    float qs_f0 = scales[scales_idx][0];
-    float qs_f1 = scales[scales_idx][1];
-    float qs_f2 = scales[scales_idx][2];
-    float qs_f3 = scales[scales_idx][3];
-    int nextgroup = offset_k + groupsize;
-
-    // Column result
-
-    float block_c[m_count][4] = {};
-
-    // Dequantize groups
-
-    int k = offset_k;
-
-    while (k < rows_8 && k < end_k)
-    {
-        if (k == nextgroup)
-        {
-            group++;
-            scales_idx++;
-            qs_f0 = scales[scales_idx][0];
-            qs_f1 = scales[scales_idx][1];
-            qs_f2 = scales[scales_idx][2];
-            qs_f3 = scales[scales_idx][3];
-            nextgroup += groupsize;
-        }
-
-        #pragma unroll
-        for (int j = 0; j < 4; j++)
-        {
-            int4 load_int4[2];
-            load_int4[0] = *((int4*) b_ptr); b_ptr += size_n;
-            load_int4[1] = *((int4*) b_ptr); b_ptr += size_n;
-
-            half2 dq[4][4];
-            dequant_8bit_8(load_int4[0].x, load_int4[1].x, dq[0], size_n);
-            dequant_8bit_8(load_int4[0].y, load_int4[1].y, dq[1], size_n);
-            dequant_8bit_8(load_int4[0].z, load_int4[1].z, dq[2], size_n);
-            dequant_8bit_8(load_int4[0].w, load_int4[1].w, dq[3], size_n);
-
-            for (int m = 0; m < m_count; m++)
-            {
-                block_c[m][0] = dot22_8_f(dq[0], a_ptr + m * a_stride, block_c[m][0], qs_f0);
-                block_c[m][1] = dot22_8_f(dq[1], a_ptr + m * a_stride, block_c[m][1], qs_f1);
-                block_c[m][2] = dot22_8_f(dq[2], a_ptr + m * a_stride, block_c[m][2], qs_f2);
-                block_c[m][3] = dot22_8_f(dq[3], a_ptr + m * a_stride, block_c[m][3], qs_f3);
-            }
-            a_ptr += 8;
-        }
-        k += 32;
-    }
-
-    while (k < rows_6 && k < end_k)
-    {
-        if (k == nextgroup)
-        {
-            group++;
-            scales_idx++;
-            qs_f0 = scales[scales_idx][0];
-            qs_f1 = scales[scales_idx][1];
-            qs_f2 = scales[scales_idx][2];
-            qs_f3 = scales[scales_idx][3];
-            nextgroup += groupsize;
-        }
-
-        #pragma unroll
-        for (int j = 0; j < 2; j++)
-        {
-            int4 load_int4[3];
-            load_int4[0] = *((int4*) b_ptr); b_ptr += size_n;
-            load_int4[1] = *((int4*) b_ptr); b_ptr += size_n;
-            load_int4[2] = *((int4*) b_ptr); b_ptr += size_n;
-
-            half2 dq[4][8];
-            dequant_6bit_16(load_int4[0].x, load_int4[1].x, load_int4[2].x, dq[0], size_n);
-            dequant_6bit_16(load_int4[0].y, load_int4[1].y, load_int4[2].y, dq[1], size_n);
-            dequant_6bit_16(load_int4[0].z, load_int4[1].z, load_int4[2].z, dq[2], size_n);
-            dequant_6bit_16(load_int4[0].w, load_int4[1].w, load_int4[2].w, dq[3], size_n);
-
-            for (int m = 0; m < m_count; m++)
-            {
-                block_c[m][0] = dot22_16_f(dq[0], a_ptr + m * a_stride, block_c[m][0], qs_f0);
-                block_c[m][1] = dot22_16_f(dq[1], a_ptr + m * a_stride, block_c[m][1], qs_f1);
-                block_c[m][2] = dot22_16_f(dq[2], a_ptr + m * a_stride, block_c[m][2], qs_f2);
-                block_c[m][3] = dot22_16_f(dq[3], a_ptr + m * a_stride, block_c[m][3], qs_f3);
-            }
-            a_ptr += 16;
-        }
-        k += 32;
-    }
-
-    while (k < rows_5 && k < end_k)
-    {
-        if (k == nextgroup)
-        {
-            group++;
-            scales_idx++;
-            qs_f0 = scales[scales_idx][0];
-            qs_f1 = scales[scales_idx][1];
-            qs_f2 = scales[scales_idx][2];
-            qs_f3 = scales[scales_idx][3];
-            nextgroup += groupsize;
-        }
-
-        #pragma unroll
-        for (int j = 0; j < 1; j++)
-        {
-            int4 load_int4[5];
-            load_int4[0] = *((int4*) b_ptr); b_ptr += size_n;
-            load_int4[1] = *((int4*) b_ptr); b_ptr += size_n;
-            load_int4[2] = *((int4*) b_ptr); b_ptr += size_n;
-            load_int4[3] = *((int4*) b_ptr); b_ptr += size_n;
-            load_int4[4] = *((int4*) b_ptr); b_ptr += size_n;
-
-            half2 dq[4][16];
-            dequant_5bit_32(load_int4[0].x, load_int4[1].x, load_int4[2].x, load_int4[3].x, load_int4[4].x, dq[0], size_n);
-            dequant_5bit_32(load_int4[0].y, load_int4[1].y, load_int4[2].y, load_int4[3].y, load_int4[4].y, dq[1], size_n);
-            dequant_5bit_32(load_int4[0].z, load_int4[1].z, load_int4[2].z, load_int4[3].z, load_int4[4].z, dq[2], size_n);
-            dequant_5bit_32(load_int4[0].w, load_int4[1].w, load_int4[2].w, load_int4[3].w, load_int4[4].w, dq[3], size_n);
-
-            for (int m = 0; m < m_count; m++)
-            {
-                block_c[m][0] = dot22_32_f(dq[0], a_ptr + m * a_stride, block_c[m][0], qs_f0);
-                block_c[m][1] = dot22_32_f(dq[1], a_ptr + m * a_stride, block_c[m][1], qs_f1);
-                block_c[m][2] = dot22_32_f(dq[2], a_ptr + m * a_stride, block_c[m][2], qs_f2);
-                block_c[m][3] = dot22_32_f(dq[3], a_ptr + m * a_stride, block_c[m][3], qs_f3);
-            }
-            a_ptr += 32;
-        }
-
-        k += 32;
-    }
-
-    while (k < rows_4 && k < end_k)
-    {
-        if (k == nextgroup)
-        {
-            group++;
-            scales_idx++;
-            qs_f0 = scales[scales_idx][0];
-            qs_f1 = scales[scales_idx][1];
-            qs_f2 = scales[scales_idx][2];
-            qs_f3 = scales[scales_idx][3];
-            nextgroup += groupsize;
-        }
-
-        #pragma unroll
-        for (int j = 0; j < 4; j++)
-        {
-            int4 load_int4[1];
-            load_int4[0] = *((int4*) b_ptr); b_ptr += size_n;
-
-            half2 dq[4][4];
-            dequant_4bit_8(load_int4[0].x, dq[0], size_n);
-            dequant_4bit_8(load_int4[0].y, dq[1], size_n);
-            dequant_4bit_8(load_int4[0].z, dq[2], size_n);
-            dequant_4bit_8(load_int4[0].w, dq[3], size_n);
-
-            for (int m = 0; m < m_count; m++)
-            {
-                block_c[m][0] = dot22_8_f(dq[0], a_ptr + m * a_stride, block_c[m][0], qs_f0);
-                block_c[m][1] = dot22_8_f(dq[1], a_ptr + m * a_stride, block_c[m][1], qs_f1);
-                block_c[m][2] = dot22_8_f(dq[2], a_ptr + m * a_stride, block_c[m][2], qs_f2);
-                block_c[m][3] = dot22_8_f(dq[3], a_ptr + m * a_stride, block_c[m][3], qs_f3);
-            }
-            a_ptr += 8;
-        }
-        k += 32;
-    }
-
-    while (k < rows_3 && k < end_k)
-    {
-        if (k == nextgroup)
-        {
-            group++;
-            scales_idx++;
-            qs_f0 = scales[scales_idx][0];
-            qs_f1 = scales[scales_idx][1];
-            qs_f2 = scales[scales_idx][2];
-            qs_f3 = scales[scales_idx][3];
-            nextgroup += groupsize;
-        }
-
-        #pragma unroll
-        for (int j = 0; j < 1; j++)
-        {
-            int4 load_int4[3];
-            load_int4[0] = *((int4*) b_ptr); b_ptr += size_n;
-            load_int4[1] = *((int4*) b_ptr); b_ptr += size_n;
-            load_int4[2] = *((int4*) b_ptr); b_ptr += size_n;
-
-            half2 dq[4][16];
-            dequant_3bit_32(load_int4[0].x, load_int4[1].x, load_int4[2].x, dq[0], size_n);
-            dequant_3bit_32(load_int4[0].y, load_int4[1].y, load_int4[2].y, dq[1], size_n);
-            dequant_3bit_32(load_int4[0].z, load_int4[1].z, load_int4[2].z, dq[2], size_n);
-            dequant_3bit_32(load_int4[0].w, load_int4[1].w, load_int4[2].w, dq[3], size_n);
-
-            for (int m = 0; m < m_count; m++)
-            {
-                block_c[m][0] = dot22_32_f(dq[0], a_ptr + m * a_stride, block_c[m][0], qs_f0);
-                block_c[m][1] = dot22_32_f(dq[1], a_ptr + m * a_stride, block_c[m][1], qs_f1);
-                block_c[m][2] = dot22_32_f(dq[2], a_ptr + m * a_stride, block_c[m][2], qs_f2);
-                block_c[m][3] = dot22_32_f(dq[3], a_ptr + m * a_stride, block_c[m][3], qs_f3);
-            }
-            a_ptr += 32;
-        }
-        k += 32;
-    }
-
-    while (k < rows_2 && k < end_k)
-    {
-        if (k == nextgroup)
-        {
-            group++;
-            scales_idx++;
-            qs_f0 = scales[scales_idx][0];
-            qs_f1 = scales[scales_idx][1];
-            qs_f2 = scales[scales_idx][2];
-            qs_f3 = scales[scales_idx][3];
-            nextgroup += groupsize;
-        }
-
-        #pragma unroll
-        for (int j = 0; j < 2; j++)
-        {
-            int4 load_int4[1];
-            load_int4[0] = *((int4*) b_ptr); b_ptr += size_n;
-
-            half2 dq[4][8];
-            dequant_2bit_16(load_int4[0].x, dq[0], size_n);
-            dequant_2bit_16(load_int4[0].y, dq[1], size_n);
-            dequant_2bit_16(load_int4[0].z, dq[2], size_n);
-            dequant_2bit_16(load_int4[0].w, dq[3], size_n);
-
-            for (int m = 0; m < m_count; m++)
-            {
-                block_c[m][0] = dot22_16_f(dq[0], a_ptr + m * a_stride, block_c[m][0], qs_f0);
-                block_c[m][1] = dot22_16_f(dq[1], a_ptr + m * a_stride, block_c[m][1], qs_f1);
-                block_c[m][2] = dot22_16_f(dq[2], a_ptr + m * a_stride, block_c[m][2], qs_f2);
-                block_c[m][3] = dot22_16_f(dq[3], a_ptr + m * a_stride, block_c[m][3], qs_f3);
-            }
-
-            a_ptr += 16;
-        }
-        k += 32;
-    }
-
-    // Accumulate column sums in c
-
-    for (int m = 0; m < m_count; m++)
-    {
-        half2* out = (half2*)c_.item_ptr(offset_m + m, n);
-        half2 result01 = __halves2half2(__float2half_rn(block_c[m][0]), __float2half_rn(block_c[m][1]));
-        half2 result23 = __halves2half2(__float2half_rn(block_c[m][2]), __float2half_rn(block_c[m][3]));
-        atomicAdd(out    , result01);
-        atomicAdd(out + 1, result23);
-    }
-}
-
-fp_gemm_half_q_half_kernel pick_gemm_half_q_half_kernel(bool first_block, const int m_count)
-{
-    #if BLOCK_M_SIZE_MAX >= 1
-    if (m_count == 1) return gemm_half_q_half_kernel<true, 1>;
-    #endif
-    #if BLOCK_M_SIZE_MAX >= 2
-    if (m_count == 2) return gemm_half_q_half_kernel<true, 2>;
-    #endif
-    #if BLOCK_M_SIZE_MAX >= 3
-    if (m_count == 3) return gemm_half_q_half_kernel<true, 3>;
-    #endif
-    #if BLOCK_M_SIZE_MAX >= 4
-    if (m_count == 4) return gemm_half_q_half_kernel<true, 4>;
-    #endif
-    #if BLOCK_M_SIZE_MAX >= 5
-    if (m_count == 5) return gemm_half_q_half_kernel<true, 5>;
-    #endif
-    #if BLOCK_M_SIZE_MAX >= 6
-    if (m_count == 6) return gemm_half_q_half_kernel<true, 6>;
-    #endif
-    #if BLOCK_M_SIZE_MAX >= 7
-    if (m_count == 7) return gemm_half_q_half_kernel<true, 7>;
-    #endif
-    #if BLOCK_M_SIZE_MAX >= 8
-    if (m_count == 8) return gemm_half_q_half_kernel<true, 8>;
-    #endif
-    return NULL;
-}

autogptq_extension/exllamav2/cuda/q_gemm_kernel_gptq.cuh

@@ -1,219 +0,0 @@
-#include "compat.cuh"
-
-__forceinline__ __device__ half2 dot22_8(half2(&dq)[4], const half* a_ptr, const half2 g_result)
-{
-    half2 result = {};
-    const half2* a2_ptr = (const half2*)a_ptr;
-    #pragma unroll
-    for (int i = 0; i < 4; i++) result = __hfma2(dq[i], *a2_ptr++, result);
-    return __hadd2(result, g_result);
-}
-
-__forceinline__ __device__ float dot22_8_f(half2(&dq)[4], const half* a_ptr)
-{
-    half2 result = {};
-    const half2* a2_ptr = (const half2*)a_ptr;
-    #pragma unroll
-    for (int i = 0; i < 4; i++) result = __hfma2(dq[i], *a2_ptr++, result);
-    return __half2float(__low2half(result)) + __half2float(__high2half(result));
-}
-
-typedef void (*fp_gemm_half_q_half_gptq_kernel)
-(
-    const half*,
-    const uint32_t*,
-    const uint32_t*,
-    const half*,
-    half*,
-    const int,
-    const int,
-    const int,
-    const int,
-    const int,
-    const uint16_t*,
-    const int,
-    const bool
-);
-
-template <bool first_block, int m_count>
-__global__ void gemm_half_q_half_gptq_kernel
-(
-    const half* __restrict__ a,
-    const uint32_t* __restrict__ b_q_weight,
-    const uint32_t* __restrict__ b_gptq_qzeros,
-    const half* __restrict__ b_gptq_scales,
-    half* __restrict__ c,
-    const int size_m,
-    const int size_n,
-    const int size_k,
-    const int groups,
-    const int groupsize,
-    const uint16_t* __restrict__ b_q_perm,
-    const int rows_4,
-    const bool clear
-)
-{
-    MatrixView_half a_(a, size_m, size_k);
-    MatrixView_half_rw c_(c, size_m, size_n);
-    MatrixView_q4_row b_gptq_qzeros_(b_gptq_qzeros, groups, size_n);
-    MatrixView_half b_gptq_scales_(b_gptq_scales, groups, size_n);
-
-    int t = threadIdx.x;
-
-    // Block
-
-    int offset_n = blockIdx.x * BLOCK_KN_SIZE * 4;
-    int offset_m = blockIdx.y * m_count;
-    int offset_k = blockIdx.z * BLOCK_KN_SIZE;
-
-    int end_n = min(offset_n + BLOCK_KN_SIZE * 4, size_n);
-    int end_m = min(offset_m + m_count, size_m);
-    int end_k = min(offset_k + BLOCK_KN_SIZE, size_k);
-
-    int n = offset_n + t * 4;
-
-    // Preload block_a
-
-    __shared__ half block_a[m_count][BLOCK_KN_SIZE];
-
-    if (offset_k + t < end_k)
-    {
-        for (int m = 0; m < m_count; ++m)
-        {
-            const half* a_ptr = a_.item_ptr(offset_m + m, 0);
-            half* block_a_ptr = block_a[m];
-
-            half a0;
-            if (b_q_perm) a0 = a_ptr[b_q_perm[offset_k + t]];
-            else a0 = a_ptr[offset_k + t];
-            block_a_ptr[t] = a0;
-        }
-    }
-
-    // Zero output
-
-    if (n >= size_n) return;
-
-    if (clear && blockIdx.z == 0) // && (threadIdx.x & 1) == 0)
-    {
-        for (int m = 0; m < m_count; m++)
-            *((uint64_t*)c_.item_ptr(offset_m + m, n)) = 0;
-    }
-
-    __syncthreads();
-
-    // Find initial group
-
-    int group = offset_k / groupsize;
-    int nextgroup = offset_k + groupsize;
-
-    // a, b offset
-
-    int qk = offset_k / (32 / 4);
-
-    const uint32_t* b_ptr = b_q_weight + qk * size_n + n;
-    const half* a_ptr = &block_a[0][0];
-    int a_stride = BLOCK_KN_SIZE;
-
-    // Initial group
-
-    int zeros[4];
-    float scales[4];
-    half2 z1z16[4][2];
-    half2 y1y16[4][2];
-    b_gptq_qzeros_.item4(zeros, group, n);
-    b_gptq_scales_.item4_f(scales, group, n);
-    dequant_4bit_8_prep_zero(zeros[0] + 1, z1z16[0], y1y16[0]);
-    dequant_4bit_8_prep_zero(zeros[1] + 1, z1z16[1], y1y16[1]);
-    dequant_4bit_8_prep_zero(zeros[2] + 1, z1z16[2], y1y16[2]);
-    dequant_4bit_8_prep_zero(zeros[3] + 1, z1z16[3], y1y16[3]);
-
-//    __syncthreads();
-
-    // Column result
-
-    float block_c[m_count][4] = {};
-
-    // Dequantize and multiply
-
-    int k = offset_k;
-    while (k < end_k)
-    {
-        if (k == nextgroup)
-        {
-            group++;
-            nextgroup += groupsize;
-            b_gptq_qzeros_.item4(zeros, group, n);
-            b_gptq_scales_.item4_f(scales, group, n);
-            dequant_4bit_8_prep_zero(zeros[0] + 1, z1z16[0], y1y16[0]);
-            dequant_4bit_8_prep_zero(zeros[1] + 1, z1z16[1], y1y16[1]);
-            dequant_4bit_8_prep_zero(zeros[2] + 1, z1z16[2], y1y16[2]);
-            dequant_4bit_8_prep_zero(zeros[3] + 1, z1z16[3], y1y16[3]);
-        }
-
-        #pragma unroll
-        for (int j = 0; j < 4; j++)
-        {
-            const int4* b_ptr4 = (int4*) b_ptr;
-            int4 load_int4 = *b_ptr4;
-
-            half2 dq[4][4];
-            dequant_4bit_8_gptq(load_int4.x, dq[0], z1z16[0], y1y16[0], size_n, false);
-            dequant_4bit_8_gptq(load_int4.y, dq[1], z1z16[1], y1y16[1], size_n, false);
-            dequant_4bit_8_gptq(load_int4.z, dq[2], z1z16[2], y1y16[2], size_n, false);
-            dequant_4bit_8_gptq(load_int4.w, dq[3], z1z16[3], y1y16[3], size_n, false);
-
-            #pragma unroll
-            for (int m = 0; m < m_count; m++)
-            {
-                block_c[m][0] = fma(dot22_8_f(dq[0], a_ptr + m * a_stride), scales[0], block_c[m][0]);
-                block_c[m][1] = fma(dot22_8_f(dq[1], a_ptr + m * a_stride), scales[1], block_c[m][1]);
-                block_c[m][2] = fma(dot22_8_f(dq[2], a_ptr + m * a_stride), scales[2], block_c[m][2]);
-                block_c[m][3] = fma(dot22_8_f(dq[3], a_ptr + m * a_stride), scales[3], block_c[m][3]);
-            }
-
-            b_ptr += size_n;
-            a_ptr += 8;
-        }
-
-        k += 32;
-    }
-
-    for (int m = 0; m < m_count; m++)
-    {
-        half2 *out = (half2*) c_.item_ptr(offset_m + m, n);
-        half2 result01 = __halves2half2(__float2half_rn(block_c[m][0]), __float2half_rn(block_c[m][1]));
-        half2 result23 = __halves2half2(__float2half_rn(block_c[m][2]), __float2half_rn(block_c[m][3]));
-        atomicAdd(out    , result01);
-        atomicAdd(out + 1, result23);
-    }
-}
-
-fp_gemm_half_q_half_gptq_kernel pick_gemm_half_q_half_gptq_kernel(bool first_block, const int m_count)
-{
-    #if BLOCK_M_SIZE_MAX >= 1
-    if (m_count == 1) return gemm_half_q_half_gptq_kernel<true, 1>;
-    #endif
-    #if BLOCK_M_SIZE_MAX >= 2
-    if (m_count == 2) return gemm_half_q_half_gptq_kernel<true, 2>;
-    #endif
-    #if BLOCK_M_SIZE_MAX >= 3
-    if (m_count == 3) return gemm_half_q_half_gptq_kernel<true, 3>;
-    #endif
-    #if BLOCK_M_SIZE_MAX >= 4
-    if (m_count == 4) return gemm_half_q_half_gptq_kernel<true, 4>;
-    #endif
-    #if BLOCK_M_SIZE_MAX >= 5
-    if (m_count == 5) return gemm_half_q_half_gptq_kernel<true, 5>;
-    #endif
-    #if BLOCK_M_SIZE_MAX >= 6
-    if (m_count == 6) return gemm_half_q_half_gptq_kernel<true, 6>;
-    #endif
-    #if BLOCK_M_SIZE_MAX >= 7
-    if (m_count == 7) return gemm_half_q_half_gptq_kernel<true, 7>;
-    #endif
-    #if BLOCK_M_SIZE_MAX >= 8
-    if (m_count == 8) return gemm_half_q_half_gptq_kernel<true, 8>;
-    #endif
-    return NULL;
-}

autogptq_extension/exllamav2/cuda/q_matrix.cu

@@ -1,603 +0,0 @@
-#include "q_matrix.cuh"
-#include "matrix_view.cuh"
-#include "util.cuh"
-
-#include "quant/qdq_2.cuh"
-#include "quant/qdq_3.cuh"
-#include "quant/qdq_4.cuh"
-#include "quant/qdq_5.cuh"
-#include "quant/qdq_6.cuh"
-#include "quant/qdq_8.cuh"
-
-#define BLOCK_KN_SIZE 128
-
-#define THREADS_X 32
-#define THREADS_Y 32
-
-// Shuffle quantized data on load
-
-__global__ void shuffle_kernel
-(
-    uint32_t* __restrict__ b_q_weight,
-    const int size_k,
-    const int size_n,
-    const int rows_8,
-    const int rows_6,
-    const int rows_5,
-    const int rows_4,
-    const int rows_3,
-    const int rows_2
-)
-{
-    int n = blockIdx.x * THREADS_X + threadIdx.x;
-    if (n >= size_n) return;
-    int k = 0;
-    uint32_t* b_ptr = b_q_weight + n;
-    while (k < rows_8) { shuffle_8bit_4 (b_ptr, size_n); b_ptr += 1 * size_n; k +=  4; }
-    while (k < rows_6) { shuffle_6bit_16(b_ptr, size_n); b_ptr += 3 * size_n; k += 16; }
-    while (k < rows_5) { shuffle_5bit_32(b_ptr, size_n); b_ptr += 5 * size_n; k += 32; }
-    while (k < rows_4) { shuffle_4bit_8 (b_ptr, size_n); b_ptr += 1 * size_n; k +=  8; }
-    while (k < rows_3) { shuffle_3bit_32(b_ptr, size_n); b_ptr += 3 * size_n; k += 32; }
-    while (k < rows_2) { shuffle_2bit_16(b_ptr, size_n); b_ptr += 1 * size_n; k += 16; }
-}
-
-
-// QMatrix constructor
-
-QMatrix::QMatrix
-(
-    const int _device,
-    const int _height,
-    const int _width,
-    const int _groups,
-
-    uint32_t* _q_weight,
-    uint16_t* _q_perm,
-    uint16_t* _q_invperm,
-    uint32_t* _q_scale,
-    half* _q_scale_max,
-    uint16_t* _q_groups,
-
-    uint32_t* _gptq_qzeros,
-    half* _gptq_scales,
-    uint32_t* _gptq_g_idx,
-
-    half* _temp_dq
-) :
-    device(_device),
-    height(_height),
-    width(_width),
-    groups(_groups),
-    temp_dq(_temp_dq)
-{
-    cudaSetDevice(device);
-
-    cuda_q_weight = _q_weight;
-    cuda_q_perm = _q_perm;
-    cuda_q_invperm = _q_invperm;
-    cuda_q_scale = _q_scale;
-    cuda_q_scale_max = _q_scale_max;
-    cuda_q_groups = _q_groups;
-    cuda_gptq_qzeros = _gptq_qzeros;
-    cuda_gptq_scales = _gptq_scales;
-
-    is_gptq = (_gptq_qzeros != NULL);
-
-    groupsize = 1;
-    while (groupsize * groups < height) groupsize *= 2;
-
-    // Create group map
-
-    rows_8 = 0;
-    rows_6 = 0;
-    rows_5 = 0;
-    rows_4 = 0;
-    rows_3 = 0;
-    rows_2 = 0;
-
-    if (!is_gptq)
-    {
-        uint16_t* cpu_q_groups = (uint16_t*)calloc(groups * 2, sizeof(uint16_t));
-        cudaMemcpy(cpu_q_groups, cuda_q_groups, groups * 2 * sizeof(uint16_t), cudaMemcpyDeviceToHost);
-
-        for (int i = 0; i < groups; i++)
-        {
-            int bits = cpu_q_groups[i * 2];
-            if (bits == 8) rows_8 += groupsize;
-            if (bits == 6) rows_6 += groupsize;
-            if (bits == 5) rows_5 += groupsize;
-            if (bits == 4) rows_4 += groupsize;
-            if (bits == 3) rows_3 += groupsize;
-            if (bits == 2) rows_2 += groupsize;
-        }
-
-        free(cpu_q_groups);
-
-        rows_6 += rows_8;
-        rows_5 += rows_6;
-        rows_4 += rows_5;
-        rows_3 += rows_4;
-        rows_2 += rows_3;
-    }
-    else
-    {
-        rows_4 = height;
-        rows_3 = height;
-        rows_2 = height;
-
-        if (_gptq_g_idx) make_sequential(_gptq_g_idx);
-    }
-
-    // Shuffle quantized data
-
-    dim3 blockDim, gridDim;
-    blockDim.x = THREADS_X;
-    blockDim.y = 1;
-    gridDim.x = DIVIDE(width, THREADS_X);
-    gridDim.y = 1;
-
-    shuffle_kernel<<<gridDim, blockDim>>>(cuda_q_weight, height, width, rows_8, rows_6, rows_5, rows_4, rows_3, rows_2);
-}
-
-
-// Reconstruct b[k,n] (GPTQ)
-
-__global__ void reconstruct_gptq_kernel
-(
-    const uint32_t* __restrict__ b_q_weight,
-    const uint16_t* __restrict__ b_q_perm,
-    const uint32_t* __restrict__ b_gptq_qzeros,
-    const half* __restrict__ b_gptq_scales,
-    //const uint16_t* __restrict__ b_q_groups,
-    const int size_k,
-    const int size_n,
-    const int groupsize,
-    const int groups,
-    half* __restrict__ b,
-    const int rows_4
-)
-{
-    MatrixView_half_rw b_(b, size_k, size_n);
-    MatrixView_q4_row b_gptq_qzeros_(b_gptq_qzeros, groups, size_n);
-    MatrixView_half b_gptq_scales_(b_gptq_scales, groups, size_n);
-
-    int offset_k = BLOCK_KN_SIZE * blockIdx.y;
-    int offset_n = BLOCK_KN_SIZE * blockIdx.x * 4;
-
-    int end_k = min(offset_k + BLOCK_KN_SIZE, size_k);
-
-    // Preload remapping table
-
-    __shared__ uint16_t perm[BLOCK_KN_SIZE];
-    int t = threadIdx.x;
-
-    if (b_q_perm)
-    {
-        if (offset_k + t < size_k)
-            perm[t] = b_q_perm[offset_k + t];
-    }
-
-    // Column
-
-    int n = offset_n + t * 4;
-    if (n >= size_n) return;
-
-    // Find initial group
-
-    int group = offset_k / groupsize;
-    int nextgroup = offset_k + groupsize;
-
-    // b offset
-
-    int qk = offset_k / (32 / 4);
-
-    const uint32_t* b_ptr = b_q_weight + qk * size_n + n;
-
-    // Initial zeros/scale
-
-    int zeros[4];
-    half2 scales[4];
-    half2 z1z16[4][2];
-    half2 y1y16[4][2];
-    b_gptq_qzeros_.item4(zeros, group, n);
-    b_gptq_scales_.item4_h2(scales, group, n);
-    dequant_4bit_8_prep_zero(zeros[0] + 1, z1z16[0], y1y16[0]);
-    dequant_4bit_8_prep_zero(zeros[1] + 1, z1z16[1], y1y16[1]);
-    dequant_4bit_8_prep_zero(zeros[2] + 1, z1z16[2], y1y16[2]);
-    dequant_4bit_8_prep_zero(zeros[3] + 1, z1z16[3], y1y16[3]);
-
-    __syncthreads();
-
-    int k = offset_k;
-    int lk = 0;
-
-    while (k < end_k)
-    {
-        if (k == nextgroup)
-        {
-            group++;
-            nextgroup += groupsize;
-            b_gptq_qzeros_.item4(zeros, group, n);
-            b_gptq_scales_.item4_h2(scales, group, n);
-            dequant_4bit_8_prep_zero(zeros[0] + 1, z1z16[0], y1y16[0]);
-            dequant_4bit_8_prep_zero(zeros[1] + 1, z1z16[1], y1y16[1]);
-            dequant_4bit_8_prep_zero(zeros[2] + 1, z1z16[2], y1y16[2]);
-            dequant_4bit_8_prep_zero(zeros[3] + 1, z1z16[3], y1y16[3]);
-        }
-
-        for (int p = 0; p < 4; p++)
-        {
-            half2 dq[4][4];
-            const int4* b_ptr4 = (int4*) b_ptr;
-            int4 load_int4 = *b_ptr4;
-
-            dequant_4bit_8_gptq(load_int4.x, dq[0], z1z16[0], y1y16[0], size_n, false);
-            dequant_4bit_8_gptq(load_int4.y, dq[1], z1z16[1], y1y16[1], size_n, false);
-            dequant_4bit_8_gptq(load_int4.z, dq[2], z1z16[2], y1y16[2], size_n, false);
-            dequant_4bit_8_gptq(load_int4.w, dq[3], z1z16[3], y1y16[3], size_n, false);
-
-            b_ptr += size_n;
-            //half* dqh = (half*)dq;
-            if (b_q_perm)
-            {
-                for (int j = 0; j < 4; j++)
-                {
-                    for (int v = 0; v < 4; v++) dq[v][j] = __hmul2(scales[v], dq[v][j]);
-                    b_.set4(perm[lk++], n, __low2half(dq[0][j]), __low2half(dq[1][j]), __low2half(dq[2][j]), __low2half(dq[3][j]));
-                    b_.set4(perm[lk++], n, __high2half(dq[0][j]), __high2half(dq[1][j]), __high2half(dq[2][j]), __high2half(dq[3][j]));
-                }
-            }
-            else
-            {
-                for (int j = 0; j < 4; j++)
-                {
-                    for (int v = 0; v < 4; v++) dq[v][j] = __hmul2(scales[v], dq[v][j]);
-                    b_.set4(offset_k + lk++, n, __low2half(dq[0][j]), __low2half(dq[1][j]), __low2half(dq[2][j]), __low2half(dq[3][j]));
-                    b_.set4(offset_k + lk++, n, __high2half(dq[0][j]), __high2half(dq[1][j]), __high2half(dq[2][j]), __high2half(dq[3][j]));
-                }
-            }
-        }
-        k += 32;
-    }
-}
-
-
-// Reconstruct b[k,n]
-
-__global__ void reconstruct_kernel
-(
-    const uint32_t* __restrict__ b_q_weight,
-    const uint16_t* __restrict__ b_q_perm,
-    const uint32_t* __restrict__ b_q_scale,
-    const half* __restrict__ b_q_scale_max,
-    //const uint16_t* __restrict__ b_q_groups,
-    const int size_k,
-    const int size_n,
-    const int groupsize,
-    const int groups,
-    half* __restrict__ b,
-    const int rows_8,
-    const int rows_6,
-    const int rows_5,
-    const int rows_4,
-    const int rows_3,
-    const int rows_2
-)
-{
-    MatrixView_half_rw b_(b, size_k, size_n);
-    MatrixView_q4_row b_q_scale_(b_q_scale, groups, size_n);
-
-    int offset_k = BLOCK_KN_SIZE * blockIdx.y;
-    int offset_n = BLOCK_KN_SIZE * blockIdx.x;
-
-    // Preload remapping table
-
-    int t = threadIdx.x;
-    __shared__ uint16_t perm[BLOCK_KN_SIZE];
-    if (offset_k + t < size_k)
-        perm[t] = b_q_perm[offset_k + t];
-
-    // Column
-
-    int n = offset_n + t;
-    if (n >= size_n) return;
-
-    // Find initial group
-
-    int group = offset_k / groupsize;
-
-    int pre_rows_8 = min(rows_8, offset_k);
-    int pre_rows_6 = offset_k > rows_8 ? min(rows_6, offset_k) - rows_8 : 0;
-    int pre_rows_5 = offset_k > rows_6 ? min(rows_5, offset_k) - rows_6 : 0;
-    int pre_rows_4 = offset_k > rows_5 ? min(rows_4, offset_k) - rows_5 : 0;
-    int pre_rows_3 = offset_k > rows_4 ? min(rows_3, offset_k) - rows_4 : 0;
-    int pre_rows_2 = offset_k > rows_3 ? min(rows_2, offset_k) - rows_3 : 0;
-    int qk = 0;
-    qk += pre_rows_8 / 32 * 8;
-    qk += pre_rows_6 / 32 * 6;
-    qk += pre_rows_5 / 32 * 5;
-    qk += pre_rows_4 / 32 * 4;
-    qk += pre_rows_3 / 32 * 3;
-    qk += pre_rows_2 / 32 * 2;
-
-    const uint32_t* b_ptr = b_q_weight + qk * size_n + n;
-
-    half qs_h = dq_scale(b_q_scale_.item(group, n), b_q_scale_max[group]);
-    half2 qs_h2 = __halves2half2(qs_h, qs_h);
-    int nextgroup = offset_k + groupsize;
-
-    int end_k = min(offset_k + BLOCK_KN_SIZE, size_k);
-    int k = offset_k;
-    int lk = 0;
-
-    __syncthreads();
-
-    while (k < rows_8 && k < end_k)
-    {
-        if (k == nextgroup) { group++; qs_h = dq_scale(b_q_scale_.item(group, n), b_q_scale_max[group]); nextgroup += groupsize; qs_h2 = __halves2half2(qs_h, qs_h); }
-        for (int p = 0; p < 4; p++)
-        {
-            half2 dq[4];
-            uint32_t q_0 = *b_ptr; b_ptr += size_n;
-            uint32_t q_1 = *b_ptr; b_ptr += size_n;
-            dequant_8bit_8(q_0, q_1, dq, size_n);
-            for (int j = 0; j < 4; j++) dq[j] = __hmul2(dq[j], qs_h2);
-            half* dqh = (half*) dq;
-            for (int j = 0; j < 8; j++) b_.set(perm[lk++], n, dqh[j]);
-        }
-        k += 32;
-    }
-
-    while (k < rows_6 && k < end_k)
-    {
-        if (k == nextgroup) { group++; qs_h = dq_scale(b_q_scale_.item(group, n), b_q_scale_max[group]); nextgroup += groupsize; qs_h2 = __halves2half2(qs_h, qs_h); }
-        for (int p = 0; p < 2; p++)
-        {
-            half2 dq[8];
-            uint32_t q_0 = *b_ptr; b_ptr += size_n;
-            uint32_t q_1 = *b_ptr; b_ptr += size_n;
-            uint32_t q_2 = *b_ptr; b_ptr += size_n;
-            dequant_6bit_16(q_0, q_1, q_2, dq, size_n);
-            for (int j = 0; j < 8; j++) dq[j] = __hmul2(dq[j], qs_h2);
-            half* dqh = (half*) dq;
-            for (int j = 0; j < 16; j++) b_.set(perm[lk++], n, dqh[j]);
-        }
-        k += 32;
-    }
-
-    while (k < rows_5 && k < end_k)
-    {
-        if (k == nextgroup) { group++; qs_h = dq_scale(b_q_scale_.item(group, n), b_q_scale_max[group]); nextgroup += groupsize; qs_h2 = __halves2half2(qs_h, qs_h); }
-        for (int p = 0; p < 1; p++)
-        {
-            half2 dq[16];
-            uint32_t q_0 = *b_ptr; b_ptr += size_n;
-            uint32_t q_1 = *b_ptr; b_ptr += size_n;
-            uint32_t q_2 = *b_ptr; b_ptr += size_n;
-            uint32_t q_3 = *b_ptr; b_ptr += size_n;
-            uint32_t q_4 = *b_ptr; b_ptr += size_n;
-            dequant_5bit_32(q_0, q_1, q_2, q_3, q_4, dq, size_n);
-            for (int j = 0; j < 16; j++) dq[j] = __hmul2(dq[j], qs_h2);
-            half* dqh = (half*) dq;
-            for (int j = 0; j < 32; j++) b_.set(perm[lk++], n, dqh[j]);
-        }
-        k += 32;
-    }
-
-    while (k < rows_4 && k < end_k)
-    {
-        if (k == nextgroup) { group++; qs_h = dq_scale(b_q_scale_.item(group, n), b_q_scale_max[group]); nextgroup += groupsize; qs_h2 = __halves2half2(qs_h, qs_h); }
-        for (int p = 0; p < 4; p++)
-        {
-            half2 dq[4];
-            uint32_t q_0 = *b_ptr; b_ptr += size_n;
-            dequant_4bit_8(q_0, dq, size_n);
-            for (int j = 0; j < 4; j++) dq[j] = __hmul2(dq[j], qs_h2);
-            half* dqh = (half*) dq;
-            for (int j = 0; j < 8; j++) b_.set(perm[lk++], n, dqh[j]);
-        }
-        k += 32;
-    }
-
-    while (k < rows_3 && k < end_k)
-    {
-        if (k == nextgroup) { group++; qs_h = dq_scale(b_q_scale_.item(group, n), b_q_scale_max[group]); nextgroup += groupsize; qs_h2 = __halves2half2(qs_h, qs_h); }
-        for (int p = 0; p < 1; p++)
-        {
-            half2 dq[16];
-            uint32_t q_0 = *b_ptr; b_ptr += size_n;
-            uint32_t q_1 = *b_ptr; b_ptr += size_n;
-            uint32_t q_2 = *b_ptr; b_ptr += size_n;
-            dequant_3bit_32(q_0, q_1, q_2, dq, size_n);
-            for (int j = 0; j < 16; j++) dq[j] = __hmul2(dq[j], qs_h2);
-            half* dqh = (half*) dq;
-            for (int j = 0; j < 32; j++) b_.set(perm[lk++], n, dqh[j]);
-        }
-        k += 32;
-    }
-
-    while (k < rows_2 && k < end_k)
-    {
-        if (k == nextgroup) { group++; qs_h = dq_scale(b_q_scale_.item(group, n), b_q_scale_max[group]); nextgroup += groupsize; qs_h2 = __halves2half2(qs_h, qs_h); }
-        for (int p = 0; p < 2; p++)
-        {
-            half2 dq[8];
-            uint32_t q_0 = *b_ptr; b_ptr += size_n;
-            dequant_2bit_16(q_0, dq, size_n);
-            for (int j = 0; j < 8; j++) dq[j] = __hmul2(dq[j], qs_h2);
-            half* dqh = (half*) dq;
-            for (int j = 0; j < 16; j++) b_.set(perm[lk++], n, dqh[j]);
-        }
-        k += 32;
-    }
-}
-
-void QMatrix::reconstruct(half* out)
-{
-    dim3 blockDim, gridDim;
-    blockDim.x = BLOCK_KN_SIZE;
-    blockDim.y = 1;
-    gridDim.x = DIVIDE(width, BLOCK_KN_SIZE);
-    gridDim.y = DIVIDE(height, BLOCK_KN_SIZE);
-
-    if (!is_gptq)
-    {
-        reconstruct_kernel<<<gridDim, blockDim>>>
-        (
-            cuda_q_weight,
-            cuda_q_perm,
-            cuda_q_scale,
-            cuda_q_scale_max,
-            //cuda_q_groups,
-            height,
-            width,
-            groupsize,
-            groups,
-            out,
-            rows_8,
-            rows_6,
-            rows_5,
-            rows_4,
-            rows_3,
-            rows_2
-        );
-    }
-    else
-    {
-        reconstruct_gptq_kernel<<<gridDim, blockDim>>>
-        (
-            cuda_q_weight,
-            cuda_q_perm,
-            cuda_gptq_qzeros,
-            cuda_gptq_scales,
-            //const uint16_t* __restrict__ b_q_groups,
-            height,
-            width,
-            groupsize,
-            groups,
-            out,
-            rows_4
-        );
-    }
-}
-
-__global__ void make_sequential_kernel
-(
-    const uint32_t* __restrict__ w,
-    uint32_t* __restrict__ w_new,
-    const uint16_t* __restrict__ q_perm,
-    const int w_height,
-    const int w_width
-)
-{
-    const uint64_t* w2 = (uint64_t*) w;
-    uint64_t* w_new2 = (uint64_t*) w_new;
-    int w2_stride = w_width >> 1;
-
-    int w2_column = THREADS_X * blockIdx.x + threadIdx.x;
-    if (w2_column >= w2_stride) return;
-
-    int w_new2_row = blockIdx.y;
-
-    int q_perm_idx = w_new2_row << 3;
-
-    uint64_t dst = 0;
-
-    #pragma unroll
-    for (int i = 0; i < 8; i++)
-    {
-        int source_row = q_perm[q_perm_idx++];
-
-        int w2_row = source_row >> 3;
-        int w2_subrow = source_row & 0x07;
-        int w2_row_shift = w2_subrow << 2;
-        int wnew2_row_shift = i << 2;
-
-        uint64_t src = w2[w2_row * w2_stride + w2_column];
-        src >>= w2_row_shift;
-        src &= 0x0000000f0000000f;
-        src <<= wnew2_row_shift;
-        dst |= src;
-    }
-
-    w_new2[w_new2_row * w2_stride + w2_column] = dst;
-}
-
-void QMatrix::make_sequential(const uint32_t* cpu_g_idx)
-{
-    uint32_t* cuda_new_qweight = NULL;
-    cudaMalloc(&cuda_new_qweight, height / 8 * width * sizeof(uint32_t));
-
-    uint32_t* cpu_g_idx_map = (uint32_t*) calloc(groups, sizeof(uint32_t));
-    uint32_t* cpu_x_map = (uint32_t*) malloc(height * sizeof(uint32_t));
-    uint32_t* cpu_x_map_inv = (uint32_t*) malloc(height * sizeof(uint32_t));
-
-    // Group histogram
-
-    for (int i = 0; i < height; i++) cpu_g_idx_map[cpu_g_idx[i]]++;
-
-    // Group map
-
-    for (int i = 0, acc = 0; i < groups; i++)
-    {
-        short tmp = cpu_g_idx_map[i];
-        cpu_g_idx_map[i] = acc;
-        acc += tmp;
-    }
-
-    // X map (inverse)
-
-    for (int row = 0; row < height; row++)
-    {
-        uint32_t target_group = cpu_g_idx[row];
-        uint32_t target_row = cpu_g_idx_map[target_group];
-        cpu_g_idx_map[target_group]++;
-        cpu_x_map_inv[row] = target_row;
-    }
-
-    // X map
-
-    for (int row = 0; row < height; row++) cpu_x_map[cpu_x_map_inv[row]] = row;
-
-    // Reduce to uint16_t
-
-    uint16_t* cpu_x_map16 = (uint16_t*)cpu_x_map;
-    uint16_t* cpu_x_map_inv16 = (uint16_t*)cpu_x_map_inv;
-    for (int row = 0; row < height; row++) cpu_x_map16[row] = (uint16_t) cpu_x_map[row];
-    for (int row = 0; row < height; row++) cpu_x_map_inv16[row] = (uint16_t) cpu_x_map_inv[row];
-
-    // Move to CUDA
-
-    cudaMemcpyAsync(cuda_q_perm, cpu_x_map16, height * sizeof(uint16_t), cudaMemcpyHostToDevice);
-    cudaMemcpyAsync(cuda_q_invperm, cpu_x_map_inv16, height * sizeof(uint16_t), cudaMemcpyHostToDevice);
-
-    // Rearrange rows in w
-
-    dim3 blockDim, gridDim;
-    blockDim.x = THREADS_X;
-    blockDim.y = 1;
-    gridDim.x = DIVIDE(width, THREADS_X);
-    gridDim.y = height / 8;
-
-    make_sequential_kernel<<<gridDim, blockDim>>>
-    (
-        cuda_q_weight,
-        cuda_new_qweight,
-        cuda_q_perm,
-        height / 8,
-        width
-    );
-
-    // Replace qweights
-
-    cudaMemcpyAsync(cuda_q_weight, cuda_new_qweight, height / 8 * width * sizeof(uint32_t), cudaMemcpyDeviceToDevice);
-
-    // Cleanup
-
-    cudaDeviceSynchronize();
-
-    cudaFree(cuda_new_qweight);
-    free(cpu_g_idx_map);
-    free(cpu_x_map);
-    free(cpu_x_map_inv);
-}

autogptq_extension/exllamav2/cuda/q_matrix.cuh

@@ -1,71 +0,0 @@
-#ifndef _q_matrix_cuh
-#define _q_matrix_cuh
-
-#include <cuda_runtime.h>
-#include <cuda_fp16.h>
-#include <cstdint>
-#include <cstdio>
-
-#define MAX_SUPERGROUPS 16
-
-class QMatrix
-{
-public:
-
-    int device;
-    bool is_gptq;
-
-    int height;
-    int width;
-    int groups;
-    int groupsize;
-
-    int rows_8;
-    int rows_6;
-    int rows_5;
-    int rows_4;
-    int rows_3;
-    int rows_2;
-
-    uint32_t* cuda_q_weight = NULL;
-    uint16_t* cuda_q_perm = NULL;
-    uint16_t* cuda_q_invperm = NULL;
-    uint32_t* cuda_q_scale = NULL;
-    half* cuda_q_scale_max = NULL;
-    uint16_t* cuda_q_groups = NULL;
-    uint32_t* cuda_gptq_qzeros = NULL;
-    half* cuda_gptq_scales = NULL;
-
-    half* temp_dq;
-
-    QMatrix
-    (
-        const int _device,
-        const int _height,
-        const int _width,
-        const int _groups,
-
-        uint32_t* _q_weight,
-        uint16_t* _q_perm,
-        uint16_t* _q_invperm,
-        uint32_t* _q_scale,
-        half* _q_scale_max,
-        uint16_t* _q_groups,
-
-        uint32_t* _gptq_qzeros,
-        half* _gptq_scales,
-        uint32_t* _gptq_g_idx,
-
-        half* _temp_dq
-    );
-
-    ~QMatrix();
-
-    void reconstruct(half* out);
-    void make_sequential(const uint32_t* cpu_g_idx);
-
-private:
-
-};
-
-#endif

autogptq_extension/exllamav2/cuda/quant/qdq_2.cuh

@@ -1,103 +0,0 @@
-#ifndef _qdq_2_cuh
-#define _qdq_2_cuh
-
-#include "qdq_util.cuh"
-#include "../../config.h"
-
-#if QMODE_2BIT == 1
-
-// Permutation:
-//
-// ffddbb99 77553311  eeccaa88 66442200
-
-__forceinline__ __device__ void shuffle_2bit_16
-(
-    uint32_t* q,
-    int stride
-)
-{
-    uint32_t qa = q[0];
-    uint32_t qb = 0;
-
-    #pragma unroll
-    for (int i = 0; i < 8; i++)
-    {
-        uint32_t qa0 = qa & 0x03;
-        uint32_t qa1 = (qa & 0x0c) >> 2;
-        qa >>= 4;
-        qb |= (qa1 << (i * 2 + 16));
-        qb |= (qa0 << (i * 2));
-    }
-    q[0] = qb;
-}
-
-__forceinline__ __device__ void dequant_2bit_16
-(
-    const uint32_t q_0,
-    half2 (&dq)[8],
-    int stride
-)
-{
-    const uint32_t c0 = 0x64006400;
-    const half y4_  = __float2half_rn(1.0f /  4.0f);
-    const half y16_ = __float2half_rn(1.0f / 16.0f);
-    const half y64_ = __float2half_rn(1.0f / 64.0f);
-    const half2 y4  = __halves2half2(y4_,  y4_);
-    const half2 y16 = __halves2half2(y16_, y16_);
-    const half2 y64 = __halves2half2(y64_, y64_);
-    const half z1_  = __float2half_rn(-1024.0f         - 2.0f);
-    const half z4_  = __float2half_rn(-1024.0f /  4.0f - 2.0f);
-    const half z16_ = __float2half_rn(-1024.0f / 16.0f - 2.0f);
-    const half z64_ = __float2half_rn(-1024.0f / 64.0f - 2.0f);
-    const half2 z1  = __halves2half2(z1_,  z1_);
-    const half2 z4  = __halves2half2(z4_,  z4_);
-    const half2 z16 = __halves2half2(z16_, z16_);
-    const half2 z64 = __halves2half2(z64_, z64_);
-
-    uint32_t qa = q_0;
-    half2_uint32 q0((qa & 0x00030003) | c0); // half2(q[ 0], q[ 1])      + 1024
-    half2_uint32 q1((qa & 0x000c000c) | c0); // half2(q[ 2], q[ 3]) *  4 + 1024
-    half2_uint32 q2((qa & 0x00300030) | c0); // half2(q[ 4], q[ 5]) * 16 + 1024
-    half2_uint32 q3((qa & 0x00c000c0) | c0); // half2(q[ 6], q[ 7]) * 64 + 1024
-    qa >>= 8;
-    half2_uint32 q4((qa & 0x00030003) | c0); // half2(q[ 8], q[ 8])      + 1024
-    half2_uint32 q5((qa & 0x000c000c) | c0); // half2(q[10], q[11]) *  4 + 1024
-    half2_uint32 q6((qa & 0x00300030) | c0); // half2(q[12], q[13]) * 16 + 1024
-    half2_uint32 q7((qa & 0x00c000c0) | c0); // half2(q[14], q[15]) * 64 + 1024
-
-    dq[0] = __hadd2(q0.as_half2, z1);
-    dq[1] = __hfma2(q1.as_half2, y4,  z4);
-    dq[2] = __hfma2(q2.as_half2, y16, z16);
-    dq[3] = __hfma2(q3.as_half2, y64, z64);
-    dq[4] = __hadd2(q4.as_half2, z1);
-    dq[5] = __hfma2(q5.as_half2, y4,  z4);
-    dq[6] = __hfma2(q6.as_half2, y16, z16);
-    dq[7] = __hfma2(q7.as_half2, y64, z64);
-}
-
-#else
-
-__forceinline__ __device__ void shuffle_2bit_16
-(
-    uint32_t* q,
-    int stride
-)
-{
-}
-
-__forceinline__ __device__ void dequant_2bit_16
-(
-    const uint32_t q_0,
-    half2 (&dq)[8],
-    int stride
-)
-{
-    half dqh[16];
-    for (int i = 0; i < 16; i++) dqh[i] = dq_ns(exb(q_0, i * 2, 0x03), 2);
-
-    for (int i = 0; i < 8; i++) dq[i] = __halves2half2(dqh[i * 2], dqh[i * 2 + 1]);
-}
-
-#endif
-
-#endif

autogptq_extension/exllamav2/cuda/quant/qdq_3.cuh

@@ -1,169 +0,0 @@
-#ifndef _qdq_3_cuh
-#define _qdq_3_cuh
-
-#include "qdq_util.cuh"
-#include "../../config.h"
-
-#if QMODE_3BIT == 1
-
-// Permutation:
-//
-// v9997775 55333111  u8886664 44222000  (u, v lsb)
-// vjjjhhhf ffdddbbb  uiiiggge eecccaaa
-// vtttrrrp ppnnnlll  usssqqqo oommmkkk
-
-__forceinline__ __device__ void shuffle_3bit_32
-(
-    uint32_t* q,
-    int stride
-)
-{
-    uint32_t qa = q[0 * stride];
-    uint32_t qb = q[1 * stride];
-    uint32_t qc = q[2 * stride];
-
-    // qa: aa999888 77766655  54443332 22111000
-    // qb: lkkkjjji iihhhggg  fffeeedd dcccbbba
-    // qc: vvvuuutt tsssrrrq  qqpppooo nnnmmmll
-
-    uint32_t qd = qc >> 26;
-    qc <<= 4;
-    qc |= qb >> 28;
-    qb <<= 2;
-    qb |= qa >> 30;
-
-    // qa: ..999888 77766655  54443332 22111000
-    // qb: ..jjjiii hhhgggff  feeedddc ccbbbaaa
-    // qc: ..tttsss rrrqqqpp  pooonnnm mmlllkkk
-    // qd:                               vvvuuu
-
-    uint32_t za = 0;
-    uint32_t zb = 0;
-    uint32_t zc = 0;
-
-    for (int i = 0; i < 5; i++) { uint32_t t0 = qa & 0x07; uint32_t t1 = (qa & 0x38) >> 3; qa >>= 6; za |= (t0 << (i * 3)); za |= (t1 << (i * 3 + 16)); }
-    for (int i = 0; i < 5; i++) { uint32_t t0 = qb & 0x07; uint32_t t1 = (qb & 0x38) >> 3; qb >>= 6; zb |= (t0 << (i * 3)); zb |= (t1 << (i * 3 + 16)); }
-    for (int i = 0; i < 5; i++) { uint32_t t0 = qc & 0x07; uint32_t t1 = (qc & 0x38) >> 3; qc >>= 6; zc |= (t0 << (i * 3)); zc |= (t1 << (i * 3 + 16)); }
-
-    // za:  9997775 55333111   8886664 44222000
-    // zb:  jjjhhhf ffdddbbb   iiiggge eecccaaa
-    // zc:  tttrrrp ppnnnlll   sssqqqo oommmkkk
-    // qd:                               vvvuuu
-
-    za |= ((qd & 0x01) >> 0) << 15;
-    zb |= ((qd & 0x02) >> 1) << 15;
-    zc |= ((qd & 0x04) >> 2) << 15;
-    za |= ((qd & 0x08) >> 3) << 31;
-    zb |= ((qd & 0x10) >> 4) << 31;
-    zc |= ((qd & 0x20) >> 5) << 31;
-
-    // za: v9997775 55333111  u8886664 44222000  (u, v lsb)
-    // zb: vjjjhhhf ffdddbbb  uiiiggge eecccaaa
-    // zc: vtttrrrp ppnnnlll  usssqqqo oommmkkk
-
-    q[0 * stride] = za;
-    q[1 * stride] = zb;
-    q[2 * stride] = zc;
-}
-
-__forceinline__ __device__ void dequant_3bit_32
-(
-    const uint32_t q_0,
-    const uint32_t q_1,
-    const uint32_t q_2,
-    half2 (&dq)[16],
-    int stride
-)
-{
-    const uint32_t c0 = 0x64006400;
-    const half y8_  = __float2half_rn(1.0f /  8.0f);
-    const half y64_ = __float2half_rn(1.0f / 64.0f);
-    const half2 y8  = __halves2half2(y8_,  y8_);
-    const half2 y64 = __halves2half2(y64_, y64_);
-    const half z1_  = __float2half_rn(-1024.0f         - 4.0f);
-    const half z8_  = __float2half_rn(-1024.0f /  8.0f - 4.0f);
-    const half z64_ = __float2half_rn(-1024.0f / 64.0f - 4.0f);
-    const half2 z1  = __halves2half2(z1_,  z1_);
-    const half2 z8  = __halves2half2(z8_,  z8_);
-    const half2 z64 = __halves2half2(z64_, z64_);
-
-    uint32_t qa = q_0;
-    uint32_t qb = q_1;
-    uint32_t qc = q_2;
-
-    half2_uint32 q0((qa & 0x00070007) | c0); // half2(q[ 0], q[ 1])      + 1024
-    half2_uint32 q1((qa & 0x00380038) | c0); // half2(q[ 2], q[ 3]) *  8 + 1024
-    qa >>= 6;
-    half2_uint32 q2((qa & 0x00070007) | c0); // half2(q[ 4], q[ 5])      + 1024
-    half2_uint32 q3((qa & 0x00380038) | c0); // half2(q[ 6], q[ 7]) *  8 + 1024
-    half2_uint32 q4((qa & 0x01c001c0) | c0); // half2(q[ 8], q[ 9]) * 64 + 1024
-    qa >>= 9;
-    qa &= 0x00010001;
-    half2_uint32 q5((qb & 0x00070007) | c0); // half2(q[10], q[11])      + 1024
-    half2_uint32 q6((qb & 0x00380038) | c0); // half2(q[12], q[13]) *  8 + 1024
-    qb >>= 6;
-    half2_uint32 q7((qb & 0x00070007) | c0); // half2(q[14], q[15])      + 1024
-    half2_uint32 q8((qb & 0x00380038) | c0); // half2(q[16], q[17]) *  8 + 1024
-    half2_uint32 q9((qb & 0x01c001c0) | c0); // half2(q[18], q[19]) * 64 + 1024
-    qb >>= 8;
-    qb &= 0x00020002;
-    half2_uint32 q10((qc & 0x00070007) | c0); // half2(q[20], q[21])      + 1024
-    half2_uint32 q11((qc & 0x00380038) | c0); // half2(q[22], q[23]) *  8 + 1024
-    qc >>= 6;
-    half2_uint32 q12((qc & 0x00070007) | c0); // half2(q[24], q[25])      + 1024
-    half2_uint32 q13((qc & 0x00380038) | c0); // half2(q[26], q[27]) *  8 + 1024
-    half2_uint32 q14((qc & 0x01c001c0) | c0); // half2(q[28], q[29]) * 64 + 1024
-    qc >>= 7;
-    qc &= 0x00040004;
-    half2_uint32 q15((qa | qb | qc) | c0);
-
-    dq[ 0] = __hadd2( q0.as_half2, z1);
-    dq[ 1] = __hfma2( q1.as_half2, y8,  z8);
-    dq[ 2] = __hadd2( q2.as_half2, z1);
-    dq[ 3] = __hfma2( q3.as_half2, y8,  z8);
-    dq[ 4] = __hfma2( q4.as_half2, y64, z64);
-    dq[ 5] = __hadd2( q5.as_half2, z1);
-    dq[ 6] = __hfma2( q6.as_half2, y8,  z8);
-    dq[ 7] = __hadd2( q7.as_half2, z1);
-    dq[ 8] = __hfma2( q8.as_half2, y8,  z8);
-    dq[ 9] = __hfma2( q9.as_half2, y64, z64);
-    dq[10] = __hadd2(q10.as_half2, z1);
-    dq[11] = __hfma2(q11.as_half2, y8,  z8);
-    dq[12] = __hadd2(q12.as_half2, z1);
-    dq[13] = __hfma2(q13.as_half2, y8,  z8);
-    dq[14] = __hfma2(q14.as_half2, y64, z64);
-    dq[15] = __hadd2(q15.as_half2, z1);
-}
-
-#else
-
-__forceinline__ __device__ void shuffle_3bit_32
-(
-    uint32_t* q,
-    int stride
-)
-{
-}
-
-__forceinline__ __device__ void dequant_3bit_32
-(
-    const uint32_t q_0,
-    const uint32_t q_1,
-    const uint32_t q_2,
-    half2 (&dq)[16],
-    int stride
-)
-{
-    half dqh[32];
-    for (int i = 0; i < 10; i++) dqh[     i] = dq_ns(exb(     q_0, i * 3    , 0x07), 4);
-                                 dqh[10    ] = dq_ns(exb(q_1, q_0,        30, 0x07), 4);
-    for (int i = 0; i < 10; i++) dqh[11 + i] = dq_ns(exb(     q_1, i * 3 + 1, 0x07), 4);
-                                 dqh[21    ] = dq_ns(exb(q_2, q_1,        31, 0x07), 4);
-    for (int i = 0; i < 10; i++) dqh[22 + i] = dq_ns(exb(     q_2, i * 3 + 2, 0x07), 4);
-
-    for (int i = 0; i < 16; i++) dq[i] = __halves2half2(dqh[i * 2], dqh[i * 2 + 1]);
-}
-
-#endif
-
-#endif

autogptq_extension/exllamav2/cuda/quant/qdq_4.cuh

@@ -1,227 +0,0 @@
-#ifndef _qdq_4_cuh
-#define _qdq_4_cuh
-
-#include "qdq_util.cuh"
-#include "../../config.h"
-
-#if QMODE_4BIT == 1
-
-// Permutation:
-//
-// 77775555 33331111  66664444 22220000
-
-__forceinline__ __device__ void shuffle_4bit_8
-(
-    uint32_t* q,
-    int stride
-)
-{
-    uint32_t qa = q[0];
-    uint32_t qb = 0;
-
-    #pragma unroll
-    for (int i = 0; i < 4; i++)
-    {
-        uint32_t qa0 = qa & 0x0f;
-        uint32_t qa1 = (qa & 0xf0) >> 4;
-        qa >>= 8;
-        qb |= (qa1 << (i * 4 + 16));
-        qb |= (qa0 << (i * 4));
-    }
-    q[0] = qb;
-}
-
-__forceinline__ __device__ void dequant_4bit_8
-(
-    const uint32_t q_0,
-    half2 (&dq)[4],
-    int stride
-)
-{
-    const uint32_t c0 = 0x64006400;
-    const half y16_ = __float2half_rn(1.0f / 16.0f);
-    const half2 y16 = __halves2half2(y16_, y16_);
-    const half z1_  = __float2half_rn(-1024.0f         - 8.0f);
-    const half z16_ = __float2half_rn(-1024.0f / 16.0f - 8.0f);
-    const half2 z1  = __halves2half2(z1_,  z1_);
-    const half2 z16 = __halves2half2(z16_, z16_);
-
-    uint32_t qa = q_0;
-    half2_uint32 q0((qa & 0x000f000f) | c0); // half2(q[ 0], q[ 1])      + 1024
-    half2_uint32 q1((qa & 0x00f000f0) | c0); // half2(q[ 2], q[ 3]) * 16 + 1024
-    qa >>= 8;
-    half2_uint32 q2((qa & 0x000f000f) | c0); // half2(q[ 4], q[ 5])      + 1024
-    half2_uint32 q3((qa & 0x00f000f0) | c0); // half2(q[ 6], q[ 7]) * 16 + 1024
-
-    dq[0] = __hadd2(q0.as_half2, z1);
-    dq[1] = __hfma2(q1.as_half2, y16, z16);
-    dq[2] = __hadd2(q2.as_half2, z1);
-    dq[3] = __hfma2(q3.as_half2, y16, z16);
-}
-
-__forceinline__ __device__ void dequant_4bit_8_prep_zero_scale
-(
-    const uint32_t zero,
-    const half scale,
-    half2 (&z1z16)[2],
-    half2 (&y1y16)[2]
-)
-{
-    half_uint16 z1(0xe400 | zero); // half(-1024.0f - zero);
-    half z16 = __hsub(__int2half_rn(-64), __int2half_rn(zero));
-
-    half2 scale2 = __half2half2(scale);
-
-    z1z16[0] = __hmul2(scale2, __half2half2(z1.as_half));
-    z1z16[1] = __hmul2(scale2, __half2half2(z16));
-
-    const half y1 = __float2half_rn(1.0f);
-    const half y16 = __float2half_rn(1.0f / 16.0f);
-
-    y1y16[0] = __hmul2(scale2, __half2half2(y1));
-    y1y16[1] = __hmul2(scale2, __half2half2(y16));
-}
-
-__forceinline__ __device__ void dequant_4bit_8_prep_zero
-(
-    const uint32_t zero,
-    half2(&z1z16)[2],
-    half2(&y1y16)[2]
-)
-{
-    half_uint16 z1(0xe400 | zero); // half(-1024.0f - zero);
-    half z16 = __hsub(__int2half_rn(-64), __int2half_rn(zero));
-
-    z1z16[0] = __half2half2(z1.as_half);
-    z1z16[1] = __half2half2(z16);
-
-    const half y1 = __float2half_rn(1.0f);
-    const half y16 = __float2half_rn(1.0f / 16.0f);
-
-    y1y16[0] = __half2half2(y1);
-    y1y16[1] = __half2half2(y16);
-}
-
-
-__forceinline__ __device__ void dequant_4bit_8_gptq
-(
-    const uint32_t q_0,
-    half2 (&dq)[4],
-    half2 (&z1z16)[2],
-    half2 (&y1y16)[2],
-    int stride,
-    bool scaled
-)
-{
-    const uint32_t c0 = 0x64006400;
-
-    uint32_t qa = q_0;
-    half2_uint32 q0((qa & 0x000f000f) | c0); // half2( q[0]      + 1024, q[1]      + 1024 )
-    half2_uint32 q1((qa & 0x00f000f0) | c0); // half2( q[2] * 16 + 1024, q[3] * 16 + 1024 )
-    qa >>= 8;
-    half2_uint32 q2((qa & 0x000f000f) | c0); // half2( q[4]      + 1024, q[5]      + 1024 )
-    half2_uint32 q3((qa & 0x00f000f0) | c0); // half2( q[6] * 16 + 1024, q[7] * 16 + 1024 )
-
-    if (scaled)
-    {
-        dq[0] = __hfma2(q0.as_half2, y1y16[0], z1z16[0]);  // half2( q[0] * s - z * s, q[1] * s - z * s)
-        dq[1] = __hfma2(q1.as_half2, y1y16[1], z1z16[1]);  // half2( q[2] * s - z * s, q[3] * s - z * s)
-        dq[2] = __hfma2(q2.as_half2, y1y16[0], z1z16[0]);
-        dq[3] = __hfma2(q3.as_half2, y1y16[1], z1z16[1]);
-    }
-    else
-    {
-        dq[0] = __hadd2(q0.as_half2,           z1z16[0]);  // half2( q[0] - z, q[1] - z )
-        dq[1] = __hfma2(q1.as_half2, y1y16[1], z1z16[1]);  // half2( q[2] - z, q[3] - z )
-        dq[2] = __hadd2(q2.as_half2,           z1z16[0]);  // half2( q[4] - z, q[5] - z )
-        dq[3] = __hfma2(q3.as_half2, y1y16[1], z1z16[1]);  // half2( q[6] - z, q[7] - z )
-    }
-}
-
-#else
-
-__forceinline__ __device__ void shuffle_4bit_8
-(
-    uint32_t* q,
-    int stride
-)
-{
-}
-
-__forceinline__ __device__ void dequant_4bit_8
-(
-    const uint32_t q_0,
-    half2 (&dq)[4],
-    int stride
-)
-{
-    half dqh[8];
-    for (int i = 0; i < 8; i++) dqh[i] = dq_ns(exb(q_0, i * 4, 0x0f), 8);
-
-    for (int i = 0; i < 4; i++) dq[i] = __halves2half2(dqh[i * 2], dqh[i * 2 + 1]);
-}
-
-__forceinline__ __device__ void dequant_4bit_8_prep_zero_scale
-(
-    const uint32_t zero,
-    const half scale,
-    half2 (&z1)[2],
-    half2 (&y1)[2]
-)
-{
-    half z = __int2half_rn(-((int)zero));
-    z = __hmul(z, scale);
-    z1[0] = __half2half2(z);
-    y1[0] = __half2half2(scale);
-}
-
-__forceinline__ __device__ void dequant_4bit_8_prep_zero
-(
-    const uint32_t zero,
-    half2(&z1)[2],
-    half2(&y1)[2]
-)
-{
-    half z = __int2half_rn(-((int)zero));
-    z1[0] = __half2half2(z);
-}
-
-__forceinline__ __device__ void dequant_4bit_8_gptq
-(
-    const uint32_t q_0,
-    half2 (&dq)[4],
-    half2 (&z1)[2],
-    half2 (&y1)[2],
-    int stride,
-    bool scaled
-)
-{
-    half2 dqh2[8];
-
-    uint32_t qa = q_0;
-    for (int i = 0; i < 4; i++)
-    {
-        half d0 = __int2half_rn(qa & 0x0f); qa >>= 4;
-        half d1 = __int2half_rn(qa & 0x0f); qa >>= 4;
-        dqh2[i] = __halves2half2(d0, d1);
-    }
-
-    if (scaled)
-    {
-        dq[0] = __hfma2(dqh2[0], y1[0], z1[0]);
-        dq[1] = __hfma2(dqh2[1], y1[0], z1[0]);
-        dq[2] = __hfma2(dqh2[2], y1[0], z1[0]);
-        dq[3] = __hfma2(dqh2[3], y1[0], z1[0]);
-    }
-    else
-    {
-        dq[0] = __hadd2(dqh2[0], z1[0]);
-        dq[1] = __hadd2(dqh2[1], z1[0]);
-        dq[2] = __hadd2(dqh2[2], z1[0]);
-        dq[3] = __hadd2(dqh2[3], z1[0]);
-    }
-}
-
-#endif
-
-#endif

autogptq_extension/exllamav2/cuda/quant/qdq_5.cuh

@@ -1,207 +0,0 @@
-#ifndef _qdq_5_cuh
-#define _qdq_5_cuh
-
-#include "qdq_util.cuh"
-#include "../../config.h"
-
-#if QMODE_5BIT == 1
-
-// Permutation:
-//
-// v5555533 33311111  u4444422 22200000  (u, v lsb)
-// vbbbbb99 99977777  uaaaaa88 88866666
-// vhhhhhff fffddddd  ugggggee eeeccccc
-// vnnnnnll llljjjjj  ummmmmkk kkkiiiii
-// vtttttrr rrrppppp  usssssqq qqqooooo
-
-__forceinline__ __device__ void shuffle_5bit_32
-(
-    uint32_t* q,
-    int stride
-)
-{
-    uint32_t qa = q[0 * stride];
-    uint32_t qb = q[1 * stride];
-    uint32_t qc = q[2 * stride];
-    uint32_t qd = q[3 * stride];
-    uint32_t qe = q[4 * stride];
-
-    // qa: 66555554 44443333  32222211 11100000
-    // qb: ccccbbbb baaaaa99  99988888 77777666
-    // qc: jiiiiihh hhhggggg  fffffeee eedddddc
-    // qd: pppooooo nnnnnmmm  mmlllllk kkkkjjjj
-    // qe: vvvvvuuu uuttttts  ssssrrrr rqqqqqpp
-
-    uint32_t qf = qe >> 22;
-    qe <<= 8;
-    qe |= qd >> 24;
-    qd <<= 6;
-    qd |= qc >> 26;
-    qc <<= 4;
-    qc |= qb >> 28;
-    qb <<= 2;
-    qb |= qa >> 30;
-
-    // qa:   555554 44443333  32222211 11100000
-    // qb:   bbbbba aaaa9999  98888877 77766666
-    // qc:   hhhhhg ggggffff  feeeeedd dddccccc
-    // qd:   nnnnnm mmmmllll  lkkkkkjj jjjiiiii
-    // qe:   ttttts ssssrrrr  rqqqqqpp pppooooo
-    // qf:                          vv vvvuuuuu
-
-    uint32_t za = 0;
-    uint32_t zb = 0;
-    uint32_t zc = 0;
-    uint32_t zd = 0;
-    uint32_t ze = 0;
-
-    for (int i = 0; i < 3; i++) { uint32_t t0 = qa & 0x1f; uint32_t t1 = (qa & 0x3e0) >> 5; qa >>= 10; za |= (t0 << (i * 5)); za |= (t1 << (i * 5 + 16)); }
-    for (int i = 0; i < 3; i++) { uint32_t t0 = qb & 0x1f; uint32_t t1 = (qb & 0x3e0) >> 5; qb >>= 10; zb |= (t0 << (i * 5)); zb |= (t1 << (i * 5 + 16)); }
-    for (int i = 0; i < 3; i++) { uint32_t t0 = qc & 0x1f; uint32_t t1 = (qc & 0x3e0) >> 5; qc >>= 10; zc |= (t0 << (i * 5)); zc |= (t1 << (i * 5 + 16)); }
-    for (int i = 0; i < 3; i++) { uint32_t t0 = qd & 0x1f; uint32_t t1 = (qd & 0x3e0) >> 5; qd >>= 10; zd |= (t0 << (i * 5)); zd |= (t1 << (i * 5 + 16)); }
-    for (int i = 0; i < 3; i++) { uint32_t t0 = qe & 0x1f; uint32_t t1 = (qe & 0x3e0) >> 5; qe >>= 10; ze |= (t0 << (i * 5)); ze |= (t1 << (i * 5 + 16)); }
-
-    // za:  5555533 33311111   4444422 22200000
-    // zb:  bbbbb99 99977777   aaaaa88 88866666
-    // zc:  hhhhhff fffddddd   gggggee eeeccccc
-    // zd:  nnnnnll llljjjjj   mmmmmkk kkkiiiii
-    // ze:  tttttrr rrrppppp   sssssqq qqqooooo
-    // qf:                          vv vvvuuuuu
-
-    za |= ((qf & 0x001) >> 0) << 15;
-    zb |= ((qf & 0x002) >> 1) << 15;
-    zc |= ((qf & 0x004) >> 2) << 15;
-    zd |= ((qf & 0x008) >> 3) << 15;
-    ze |= ((qf & 0x010) >> 4) << 15;
-    za |= ((qf & 0x020) >> 5) << 31;
-    zb |= ((qf & 0x040) >> 6) << 31;
-    zc |= ((qf & 0x080) >> 7) << 31;
-    zd |= ((qf & 0x100) >> 8) << 31;
-    ze |= ((qf & 0x200) >> 9) << 31;
-
-    // za: v5555533 33311111  u4444422 22200000  (u, v lsb)
-    // zb: vbbbbb99 99977777  uaaaaa88 88866666
-    // zc: vhhhhhff fffddddd  ugggggee eeeccccc
-    // zd: vnnnnnll llljjjjj  ummmmmkk kkkiiiii
-    // ze: vtttttrr rrrppppp  usssssqq qqqooooo
-
-    q[0 * stride] = za;
-    q[1 * stride] = zb;
-    q[2 * stride] = zc;
-    q[3 * stride] = zd;
-    q[4 * stride] = ze;
-}
-
-__forceinline__ __device__ void dequant_5bit_32
-(
-    const uint32_t q_0,
-    const uint32_t q_1,
-    const uint32_t q_2,
-    const uint32_t q_3,
-    const uint32_t q_4,
-    half2 (&dq)[16],
-    int stride
-)
-{
-    const uint32_t c0 = 0x64006400;
-    const half y32_ = __float2half_rn(1.0f / 32.0f);
-    const half2 y32 = __halves2half2(y32_, y32_);
-    const half z1_  = __float2half_rn(-1024.0f         - 16.0f);
-    const half z32_ = __float2half_rn(-1024.0f / 32.0f - 16.0f);
-    const half2 z1  = __halves2half2(z1_,  z1_);
-    const half2 z32 = __halves2half2(z32_, z32_);
-
-    uint32_t qa = q_0;
-    uint32_t qb = q_1;
-    uint32_t qc = q_2;
-    uint32_t qd = q_3;
-    uint32_t qe = q_4;
-
-    half2_uint32 q0 ((qa & 0x001f001f) | c0); // half2(q[ 0], q[ 1])      + 1024
-    half2_uint32 q1 ((qa & 0x03e003e0) | c0); // half2(q[ 2], q[ 3]) * 32 + 1024
-    qa >>= 10;
-    half2_uint32 q2 ((qa & 0x001f001f) | c0); // half2(q[ 4], q[ 5])      + 1024
-    qa >>= 5;
-    qa &= 0x00010001;
-    half2_uint32 q3 ((qb & 0x001f001f) | c0); // half2(q[ 6], q[ 7])      + 1024
-    half2_uint32 q4 ((qb & 0x03e003e0) | c0); // half2(q[ 8], q[ 9]) * 32 + 1024
-    qb >>= 10;
-    half2_uint32 q5 ((qb & 0x001f001f) | c0); // half2(q[10], q[11])      + 1024
-    qb >>= 4;
-    qb &= 0x00020002;
-    half2_uint32 q6 ((qc & 0x001f001f) | c0); // half2(q[12], q[13])      + 1024
-    half2_uint32 q7 ((qc & 0x03e003e0) | c0); // half2(q[14], q[15]) * 32 + 1024
-    qc >>= 10;
-    half2_uint32 q8 ((qc & 0x001f001f) | c0); // half2(q[16], q[17])      + 1024
-    qc >>= 3;
-    qc &= 0x00040004;
-    half2_uint32 q9 ((qd & 0x001f001f) | c0); // half2(q[18], q[19])      + 1024
-    half2_uint32 q10((qd & 0x03e003e0) | c0); // half2(q[20], q[21]) * 32 + 1024
-    qd >>= 10;
-    half2_uint32 q11((qd & 0x001f001f) | c0); // half2(q[22], q[23])      + 1024
-    qd >>= 2;
-    qd &= 0x00080008;
-    half2_uint32 q12((qe & 0x001f001f) | c0); // half2(q[24], q[25])      + 1024
-    half2_uint32 q13((qe & 0x03e003e0) | c0); // half2(q[26], q[27]) * 32 + 1024
-    qe >>= 10;
-    half2_uint32 q14((qe & 0x001f001f) | c0); // half2(q[28], q[29])      + 1024
-    qe >>= 1;
-    qe &= 0x00100010;
-    half2_uint32 q15((qa | qb | qc | qd | qe) | c0);
-
-    dq[ 0] = __hadd2( q0.as_half2, z1);
-    dq[ 1] = __hfma2( q1.as_half2, y32, z32);
-    dq[ 2] = __hadd2( q2.as_half2, z1);
-    dq[ 3] = __hadd2( q3.as_half2, z1);
-    dq[ 4] = __hfma2( q4.as_half2, y32, z32);
-    dq[ 5] = __hadd2( q5.as_half2, z1);
-    dq[ 6] = __hadd2( q6.as_half2, z1);
-    dq[ 7] = __hfma2( q7.as_half2, y32, z32);
-    dq[ 8] = __hadd2( q8.as_half2, z1);
-    dq[ 9] = __hadd2( q9.as_half2, z1);
-    dq[10] = __hfma2(q10.as_half2, y32, z32);
-    dq[11] = __hadd2(q11.as_half2, z1);
-    dq[12] = __hadd2(q12.as_half2, z1);
-    dq[13] = __hfma2(q13.as_half2, y32, z32);
-    dq[14] = __hadd2(q14.as_half2, z1);
-    dq[15] = __hadd2(q15.as_half2, z1);
-}
-
-#else
-
-__forceinline__ __device__ void shuffle_5bit_32
-(
-    uint32_t* q,
-    int stride
-)
-{
-}
-
-__forceinline__ __device__ void dequant_5bit_32
-(
-    const uint32_t q_0,
-    const uint32_t q_1,
-    const uint32_t q_2,
-    const uint32_t q_3,
-    const uint32_t q_4,
-    half2 (&dq)[16],
-    int stride
-)
-{
-    half dqh[32];
-    for (int i = 0; i <  6; i++) dqh[     i] = dq_ns(exb(     q_0, i * 5    , 0x1f), 16);
-                                 dqh[ 6    ] = dq_ns(exb(q_1, q_0,        30, 0x1f), 16);
-    for (int i = 0; i <  5; i++) dqh[ 7 + i] = dq_ns(exb(     q_1, i * 5 + 3, 0x1f), 16);
-                                 dqh[12    ] = dq_ns(exb(q_2, q_1,        28, 0x1f), 16);
-    for (int i = 0; i <  6; i++) dqh[13 + i] = dq_ns(exb(     q_2, i * 5 + 1, 0x1f), 16);
-                                 dqh[19    ] = dq_ns(exb(q_3, q_2,        31, 0x1f), 16);
-    for (int i = 0; i <  5; i++) dqh[20 + i] = dq_ns(exb(     q_3, i * 5 + 4, 0x1f), 16);
-                                 dqh[25    ] = dq_ns(exb(q_4, q_3,        29, 0x1f), 16);
-    for (int i = 0; i <  6; i++) dqh[26 + i] = dq_ns(exb(     q_4, i * 5 + 2, 0x1f), 16);
-
-    for (int i = 0; i < 16; i++) dq[i] = __halves2half2(dqh[i * 2], dqh[i * 2 + 1]);
-}
-
-#endif
-
-#endif

autogptq_extension/exllamav2/cuda/quant/qdq_6.cuh

@@ -1,44 +0,0 @@
-#ifndef _qdq_6_cuh
-#define _qdq_6_cuh
-
-#include "qdq_util.cuh"
-#include "../../config.h"
-
-#if QMODE_6BIT == 1
-
-  // Not implemented
-
-#else
-
-__forceinline__ __device__ void shuffle_6bit_16
-(
-    uint32_t* q,
-    int stride
-)
-{
-}
-
-__forceinline__ __device__ void dequant_6bit_16
-(
-    const uint32_t q_0,
-    const uint32_t q_1,
-    const uint32_t q_2,
-    half2 (&dq)[8],
-    int stride
-)
-{
-    half dqh[16];
-    for (int i = 0; i < 5; i++) dqh[     i] = dq_ns(exb(     q_0, i * 6    , 0x3f), 32);
-                                dqh[ 5    ] = dq_ns(exb(q_1, q_0,        30, 0x3f), 32);
-    for (int i = 0; i < 4; i++) dqh[ 6 + i] = dq_ns(exb(     q_1, i * 6 + 4, 0x3f), 32);
-                                dqh[10    ] = dq_ns(exb(q_2, q_1,        28, 0x3f), 32);
-    for (int i = 0; i < 5; i++) dqh[11 + i] = dq_ns(exb(     q_2, i * 6 + 2, 0x3f), 32);
-
-    for (int i = 0; i < 8; i++) dq[i] = __halves2half2(dqh[i * 2], dqh[i * 2 + 1]);
-}
-
-#endif
-
-#endif
-
-

autogptq_extension/exllamav2/cuda/quant/qdq_8.cuh

@@ -1,38 +0,0 @@
-#ifndef _qdq_8_cuh
-#define _qdq_8_cuh
-
-#include "qdq_util.cuh"
-#include "../../config.h"
-
-#if QMODE_8BIT == 1
-
-  // Not implemented
-
-#else
-
-__forceinline__ __device__ void shuffle_8bit_4
-(
-    uint32_t* q,
-    int stride
-)
-{
-}
-
-__forceinline__ __device__ void dequant_8bit_8
-(
-    const uint32_t q_0,
-    const uint32_t q_1,
-    half2 (&dq)[4],
-    int stride
-)
-{
-    half dqh[8];
-    for (int i = 0; i < 4; i++) dqh[i    ] = dq_ns(exb(q_0, i * 8, 0xff), 128);
-    for (int i = 0; i < 4; i++) dqh[i + 4] = dq_ns(exb(q_1, i * 8, 0xff), 128);
-
-    for (int i = 0; i < 4; i++) dq[i] = __halves2half2(dqh[i * 2], dqh[i * 2 + 1]);
-}
-
-#endif
-
-#endif

autogptq_extension/exllamav2/cuda/quant/qdq_util.cuh

@@ -1,51 +0,0 @@
-#ifndef _qdq_util_cuh
-#define _qdq_util_cuh
-
-union half2_uint32
-{
-    uint32_t as_uint32;
-    half2 as_half2;
-    __device__ half2_uint32(uint32_t val) : as_uint32(val) {}
-    __device__ half2_uint32(half2 val) : as_half2(val) {}
-};
-
-union half_uint16
-{
-    uint16_t as_uint16;
-    half as_half;
-    __device__ half_uint16(uint16_t val) : as_uint16(val) {}
-    __device__ half_uint16(half val) : as_half(val) {}
-};
-
-// Max_scale premultiplied by 1/256
-
-__forceinline__ __device__ half dq_scale(const int qs, const half max_scale)
-{
-    int qs_i = qs + 1;
-    half qs_h = __int2half_rn(qs_i * qs_i);
-    qs_h = __hmul(qs_h, max_scale);
-    return qs_h;
-}
-
-__forceinline__ __device__ half dq(const int q, const int qzero, const half scale)
-{
-    return __hmul(__int2half_rn(q - qzero), scale);
-}
-
-__forceinline__ __device__ half dq_ns(const int q, const int qzero)
-{
-    //return __hsub(__int2half_rn(q), __int2half_rn(qzero));
-    return __int2half_rn(q - qzero);
-}
-
-__forceinline__ __device__ int exb(const uint32_t q, const int shift, const int mask)
-{
-    return (int)((q >> shift) & mask);
-}
-
-__forceinline__ __device__ int exb(const uint32_t q1, const uint32_t q0, const int shift, const int mask)
-{
-    return (int)(__funnelshift_rc(q0, q1, shift) & mask);
-}
-
-#endif

autogptq_extension/exllamav2/cuda/util.cuh

@@ -1,32 +0,0 @@
-
-#define DIVIDE(x, size) (((x) + (size) - 1) / (size))
-
-#define DBGS(__x) printf("%s\n", __x)
-#define DBGI(__x) printf("%s: %i\n", #__x, __x)
-#define DBGI2(__x, __y) printf("%s, %s: %i, %i\n", #__x, #__y, __x, __y)
-#define DBGI3(__x, __y, __z) printf("%s, %s, %s: %i, %i, %i\n", #__x, #__y, #__z, __x, __y, __z)
-#define DBGX(__x) printf("%s: %x\n", #__x, __x)
-#define DBGX2(__x, __y) printf("%s, %s: %x, %x\n", #__x, #__y, __x, __y)
-#define DBGX3(__x, __y, __z) printf("%s, %s, %s: %x, %x, %x\n", #__x, #__y, #__z, __x, __y, __z)
-#define DBGF(__x) printf("%s: %f\n", #__x, __x)
-#define DBGF2(__x, __y) printf("%s, %s: %f, %f\n", #__x, #__y, __x, __y)
-#define DBGF3(__x, __y, __z) printf("%s, %s, %s: %f, %f, %f\n", #__x, #__y, #__z, __x, __y, __z)
-#define DBGH(__x) printf("%s: %f\n", #__x, __half2float(__x))
-#define DBGH2(__x, __y) printf("%s, %s: %f, %f\n", #__x, #__y, __half2float(__x), __half2float(__y))
-#define DBGH3(__x, __y, __z) printf("%s, %s, %s: %f, %f, %f\n", #__x, #__y, #__z, __half2float(__x), __half2float(__y), __half2float(__z))
-
-#define DBGIH(__x, __y) printf("%s, %s: %i, %f\n", #__x, #__y, __x, __half2float(__y))
-#define DBGIH2(__x, __y, __z) printf("%s, %s, %s: %i, %f, %f\n", #__x, #__y, #__z, __x, __half2float(__y), __half2float(__z))
-
-__forceinline__ __device__ half dq_scale_(const int qs, const half max_scale)
-{
-    half qs_h = __hmul(__int2half_rn(qs + 1), __float2half_rn(1.0f / 16.0f));
-    qs_h = __hmul(qs_h, qs_h);
-    qs_h = __hmul(qs_h, max_scale);
-    return qs_h;
-}
-
-__forceinline__ __device__ float clamp(float x, float a, float b)
-{
-    return fmaxf(a, fminf(b, x));
-}

autogptq_extension/exllamav2/ext.cpp

@@ -1,134 +0,0 @@
-#include <torch/extension.h>
-#include <c10/cuda/CUDAGuard.h>
-#include <ATen/cuda/CUDAContext.h>
-#include <cuda_runtime.h>
-#include <cuda_fp16.h>
-#include <cstdint>
-#include <cstdio>
-
-#include "config.h"
-
-#include "cuda/q_matrix.cuh"
-#include "cuda/q_gemm.cuh"
-
-#include "cpp/util.h"
-
-// Some decluttering macros
-
-#define TORCH_CHECK_DTYPE(__x, __dtype) TORCH_CHECK((__x).dtype() == torch::__dtype, #__x " is incorrect datatype, must be " #__dtype)
-#define TORCH_CHECK_DTYPE_OPT(__x, __dtype) TORCH_CHECK((__x).device().is_meta() || (__x).dtype() == torch::__dtype, #__x " is incorrect datatype, must be " #__dtype)
-#define TORCH_CHECK_SHAPES(__x, __dim_x, __y, __dim_y, __scale_y) TORCH_CHECK((__x).size(__dim_x) == (__y).size(__dim_y) * __scale_y, #__x " and " #__y " have incompatible shapes")
-#define TORCH_CHECK_SHAPES_OPT(__x, __dim_x, __y, __dim_y, __scale_y) TORCH_CHECK((__x).device().is_meta() || (__x).size(__dim_x) == (__y).size(__dim_y) * __scale_y, #__x " and " #__y " have incompatible shapes")
-
-
-// Quant matrix
-
-uintptr_t make_q_matrix
-(
-    torch::Tensor q_weight,
-    torch::Tensor q_perm,
-    torch::Tensor q_invperm,
-    torch::Tensor q_scale,
-    torch::Tensor q_scale_max,
-    torch::Tensor q_groups,
-    torch::Tensor gptq_qzeros,
-    torch::Tensor gptq_scales,
-    torch::Tensor gptq_g_idx,
-    torch::Tensor temp_dq
-)
-{
-    TORCH_CHECK_DTYPE(q_weight, kInt);
-    TORCH_CHECK_DTYPE_OPT(q_perm, kShort);
-    TORCH_CHECK_DTYPE_OPT(q_invperm, kShort);
-    TORCH_CHECK_DTYPE_OPT(q_scale, kInt);
-    TORCH_CHECK_DTYPE_OPT(q_scale_max, kHalf);
-    TORCH_CHECK_DTYPE_OPT(q_groups, kShort);
-    TORCH_CHECK_DTYPE_OPT(gptq_qzeros, kInt);
-    TORCH_CHECK_DTYPE_OPT(gptq_scales, kHalf);
-    TORCH_CHECK_DTYPE_OPT(gptq_g_idx, kInt);
-
-    TORCH_CHECK_SHAPES(q_perm, 0, q_invperm, 0, 1);
-
-    int device = q_weight.device().index();
-    int width = q_weight.size(1);
-    int groups;
-    int height;
-
-    if (!q_scale.device().is_meta())
-    {
-        TORCH_CHECK_SHAPES(q_weight, 1, q_scale, 1, 8);
-        TORCH_CHECK_SHAPES(q_scale_max, 0, q_scale, 0, 1);
-        groups = q_scale.size(0);
-        height = q_invperm.size(0);
-    }
-    else
-    {
-        TORCH_CHECK_SHAPES(q_weight, 1, gptq_qzeros, 1, 8);
-        TORCH_CHECK_SHAPES(q_weight, 1, gptq_scales, 1, 1);
-        groups = gptq_qzeros.size(0);
-        height = q_weight.size(0) * 8;
-    }
-
-    TORCH_CHECK(temp_dq.size(0) >= width * height, "Insufficient size of temp_dq buffer")
-
-    QMatrix* m = new QMatrix
-    (
-        device,
-        height,
-        width,
-        groups,
-        (uint32_t*) q_weight.data_ptr(),
-        q_perm.device().is_meta() ? NULL : (uint16_t*) q_perm.data_ptr(),
-        q_invperm.device().is_meta() ? NULL : (uint16_t*) q_invperm.data_ptr(),
-        q_scale.device().is_meta() ? NULL : (uint32_t*) q_scale.data_ptr(),
-        q_scale_max.device().is_meta() ? NULL : (half*) q_scale_max.data_ptr(),
-        q_groups.device().is_meta() ? NULL : (uint16_t*) q_groups.data_ptr(),
-        gptq_qzeros.device().is_meta() ? NULL : (uint32_t*) gptq_qzeros.data_ptr(),
-        gptq_scales.device().is_meta() ? NULL : (half*) gptq_scales.data_ptr(),
-        gptq_g_idx.device().is_meta() ? NULL : (uint32_t*) gptq_g_idx.data_ptr(),
-        (half*) temp_dq.data_ptr()
-    );
-
-    return reinterpret_cast<uintptr_t> (m);
-}
-
-void gemm_half_q_half
-(
-    torch::Tensor a,
-    uintptr_t b,
-    torch::Tensor c,
-    bool force_cuda
-)
-{
-    QMatrix* qm = reinterpret_cast<QMatrix*> (b);
-
-    TORCH_CHECK_DTYPE(a, kHalf);
-    TORCH_CHECK_DTYPE(c, kHalf);
-    TORCH_CHECK_SHAPES(a, 0, c, 0, 1);
-    TORCH_CHECK(qm->height == a.size(1), "a and b have incompatible shapes")
-    TORCH_CHECK(qm->width == c.size(1), "b and c have incompatible shapes")
-
-    const at::cuda::OptionalCUDAGuard device_guard(device_of(a));
-
-    gemm_half_q_half_cuda
-    (
-        at::cuda::getCurrentCUDABlasHandle(),
-        (const half*) a.data_ptr(),
-        qm,
-        (half*) c.data_ptr(),
-        c.size(0), // m
-        c.size(1), // n
-        a.size(1), // k
-        true,
-        NULL,
-        force_cuda
-    );
-}
-
-// Bindings
-
-PYBIND11_MODULE(TORCH_EXTENSION_NAME, m)
-{
-    m.def("make_q_matrix", &make_q_matrix, "make_q_matrix");
-    m.def("gemm_half_q_half", &gemm_half_q_half, "gemm_half_q_half");
-}

autogptq_extension/qigen/forward.h

@@ -1,480 +0,0 @@
-#include<omp.h>
-#include<immintrin.h>
-#include<fstream>
-
-#define mymin(a,b) ((a)<(b)?(a):(b))
-#define mymax(a,b) ((a)>(b)?(a):(b))
-inline
-void q2gemm_gs(const float* __restrict__ input, 
-const int* __restrict__ W, 
-const float* __restrict__ scales, 
-const float* __restrict__ zeros, 
-const float* __restrict__ bias, 
- const float* __restrict__ sums, 
- float* __restrict__ output,
-const int n,
-const int m,
-const int t,
-const int nb,
-const int mb,
-const int tb,
-int ogtt,
-const int gs,
-const int cutoff){
-#pragma omp parallel num_threads(8)
-{
-int tid;
-const int mu = 16;
-const int nu = 1;
-const int tu = 32;
-const int on = n / nb;
-const int om = m / mb;
-const __m256i mask = _mm256_set1_epi32(3);
-tid = omp_get_thread_num();
-int tt = ogtt;
-if(tid >= cutoff){
-tt -= tb;
-}
-const int base_output = tid >= cutoff ?
- (tid-cutoff)*tt + (tt+tb)*cutoff: 
- tid*tt;
-const int base_W = tid >= cutoff ?
- ((tid-cutoff)*tt + (tt+tb)*cutoff)*m/16: 
- tid*tt*m/16;
-for(int j = 0; j < tt; j+=tb){
-for(int i = 0; i < on; i++) {
-for(int k = 0; k < om; k++) {
-for(int i1 = 0; i1 < nb; i1+=nu) {
-int j1 = 0;
-for(; j1 < tb-tu+1; j1+=tu) {
-for(int k1 = 0; k1 < mb; k1+=gs) {
-__m256 acc0_0 = _mm256_setzero_ps();
-__m256 acc0_8 = _mm256_setzero_ps();
-__m256 acc0_16 = _mm256_setzero_ps();
-__m256 acc0_24 = _mm256_setzero_ps();
-for(int k2 = k1; k2 < k1+gs; k2+=16)
-{
-__m256i w0 = _mm256_loadu_si256((__m256i*)&W[base_W + j*m/16 + k*mb*tb/16 + k2*tb/16 + j1+0]);
-__m256i w8 = _mm256_loadu_si256((__m256i*)&W[base_W + j*m/16 + k*mb*tb/16 + k2*tb/16 + j1+8]);
-__m256i w16 = _mm256_loadu_si256((__m256i*)&W[base_W + j*m/16 + k*mb*tb/16 + k2*tb/16 + j1+16]);
-__m256i w24 = _mm256_loadu_si256((__m256i*)&W[base_W + j*m/16 + k*mb*tb/16 + k2*tb/16 + j1+24]);
-__m256 v0_15 = _mm256_set1_ps(input[(i*om+k)*mb*nb + (k2+15)*nb + i1+0]);
-__m256 v0_14 = _mm256_set1_ps(input[(i*om+k)*mb*nb + (k2+14)*nb + i1+0]);
-__m256 v0_13 = _mm256_set1_ps(input[(i*om+k)*mb*nb + (k2+13)*nb + i1+0]);
-__m256 v0_12 = _mm256_set1_ps(input[(i*om+k)*mb*nb + (k2+12)*nb + i1+0]);
-__m256 v0_11 = _mm256_set1_ps(input[(i*om+k)*mb*nb + (k2+11)*nb + i1+0]);
-__m256 v0_10 = _mm256_set1_ps(input[(i*om+k)*mb*nb + (k2+10)*nb + i1+0]);
-__m256 v0_9 = _mm256_set1_ps(input[(i*om+k)*mb*nb + (k2+9)*nb + i1+0]);
-__m256 v0_8 = _mm256_set1_ps(input[(i*om+k)*mb*nb + (k2+8)*nb + i1+0]);
-__m256i ws0_8 = _mm256_srli_epi32(w0, 16);
-__m256i ws8_8 = _mm256_srli_epi32(w8, 16);
-__m256i ws16_8 = _mm256_srli_epi32(w16, 16);
-__m256i ws24_8 = _mm256_srli_epi32(w24, 16);
-__m256i wsa0_8= _mm256_and_si256(ws0_8, mask);
-__m256i wsa8_8= _mm256_and_si256(ws8_8, mask);
-__m256i wsa16_8= _mm256_and_si256(ws16_8, mask);
-__m256i wsa24_8= _mm256_and_si256(ws24_8, mask);
-__m256 l0_8 = _mm256_cvtepi32_ps(wsa0_8);
-__m256 l8_8 = _mm256_cvtepi32_ps(wsa8_8);
-__m256 l16_8 = _mm256_cvtepi32_ps(wsa16_8);
-__m256 l24_8 = _mm256_cvtepi32_ps(wsa24_8);
-acc0_0 = _mm256_fmadd_ps(v0_8, l0_8, acc0_0);
-acc0_8 = _mm256_fmadd_ps(v0_8, l8_8, acc0_8);
-acc0_16 = _mm256_fmadd_ps(v0_8, l16_8, acc0_16);
-acc0_24 = _mm256_fmadd_ps(v0_8, l24_8, acc0_24);
-__m256i ws0_9 = _mm256_srli_epi32(w0, 18);
-__m256i ws8_9 = _mm256_srli_epi32(w8, 18);
-__m256i ws16_9 = _mm256_srli_epi32(w16, 18);
-__m256i ws24_9 = _mm256_srli_epi32(w24, 18);
-__m256i wsa0_9= _mm256_and_si256(ws0_9, mask);
-__m256i wsa8_9= _mm256_and_si256(ws8_9, mask);
-__m256i wsa16_9= _mm256_and_si256(ws16_9, mask);
-__m256i wsa24_9= _mm256_and_si256(ws24_9, mask);
-__m256 l0_9 = _mm256_cvtepi32_ps(wsa0_9);
-__m256 l8_9 = _mm256_cvtepi32_ps(wsa8_9);
-__m256 l16_9 = _mm256_cvtepi32_ps(wsa16_9);
-__m256 l24_9 = _mm256_cvtepi32_ps(wsa24_9);
-acc0_0 = _mm256_fmadd_ps(v0_9, l0_9, acc0_0);
-acc0_8 = _mm256_fmadd_ps(v0_9, l8_9, acc0_8);
-acc0_16 = _mm256_fmadd_ps(v0_9, l16_9, acc0_16);
-acc0_24 = _mm256_fmadd_ps(v0_9, l24_9, acc0_24);
-__m256i ws0_10 = _mm256_srli_epi32(w0, 20);
-__m256i ws8_10 = _mm256_srli_epi32(w8, 20);
-__m256i ws16_10 = _mm256_srli_epi32(w16, 20);
-__m256i ws24_10 = _mm256_srli_epi32(w24, 20);
-__m256i wsa0_10= _mm256_and_si256(ws0_10, mask);
-__m256i wsa8_10= _mm256_and_si256(ws8_10, mask);
-__m256i wsa16_10= _mm256_and_si256(ws16_10, mask);
-__m256i wsa24_10= _mm256_and_si256(ws24_10, mask);
-__m256 l0_10 = _mm256_cvtepi32_ps(wsa0_10);
-__m256 l8_10 = _mm256_cvtepi32_ps(wsa8_10);
-__m256 l16_10 = _mm256_cvtepi32_ps(wsa16_10);
-__m256 l24_10 = _mm256_cvtepi32_ps(wsa24_10);
-acc0_0 = _mm256_fmadd_ps(v0_10, l0_10, acc0_0);
-acc0_8 = _mm256_fmadd_ps(v0_10, l8_10, acc0_8);
-acc0_16 = _mm256_fmadd_ps(v0_10, l16_10, acc0_16);
-acc0_24 = _mm256_fmadd_ps(v0_10, l24_10, acc0_24);
-__m256i ws0_11 = _mm256_srli_epi32(w0, 22);
-__m256i ws8_11 = _mm256_srli_epi32(w8, 22);
-__m256i ws16_11 = _mm256_srli_epi32(w16, 22);
-__m256i ws24_11 = _mm256_srli_epi32(w24, 22);
-__m256i wsa0_11= _mm256_and_si256(ws0_11, mask);
-__m256i wsa8_11= _mm256_and_si256(ws8_11, mask);
-__m256i wsa16_11= _mm256_and_si256(ws16_11, mask);
-__m256i wsa24_11= _mm256_and_si256(ws24_11, mask);
-__m256 l0_11 = _mm256_cvtepi32_ps(wsa0_11);
-__m256 l8_11 = _mm256_cvtepi32_ps(wsa8_11);
-__m256 l16_11 = _mm256_cvtepi32_ps(wsa16_11);
-__m256 l24_11 = _mm256_cvtepi32_ps(wsa24_11);
-acc0_0 = _mm256_fmadd_ps(v0_11, l0_11, acc0_0);
-acc0_8 = _mm256_fmadd_ps(v0_11, l8_11, acc0_8);
-acc0_16 = _mm256_fmadd_ps(v0_11, l16_11, acc0_16);
-acc0_24 = _mm256_fmadd_ps(v0_11, l24_11, acc0_24);
-__m256i ws0_12 = _mm256_srli_epi32(w0, 24);
-__m256i ws8_12 = _mm256_srli_epi32(w8, 24);
-__m256i ws16_12 = _mm256_srli_epi32(w16, 24);
-__m256i ws24_12 = _mm256_srli_epi32(w24, 24);
-__m256i wsa0_12= _mm256_and_si256(ws0_12, mask);
-__m256i wsa8_12= _mm256_and_si256(ws8_12, mask);
-__m256i wsa16_12= _mm256_and_si256(ws16_12, mask);
-__m256i wsa24_12= _mm256_and_si256(ws24_12, mask);
-__m256 l0_12 = _mm256_cvtepi32_ps(wsa0_12);
-__m256 l8_12 = _mm256_cvtepi32_ps(wsa8_12);
-__m256 l16_12 = _mm256_cvtepi32_ps(wsa16_12);
-__m256 l24_12 = _mm256_cvtepi32_ps(wsa24_12);
-acc0_0 = _mm256_fmadd_ps(v0_12, l0_12, acc0_0);
-acc0_8 = _mm256_fmadd_ps(v0_12, l8_12, acc0_8);
-acc0_16 = _mm256_fmadd_ps(v0_12, l16_12, acc0_16);
-acc0_24 = _mm256_fmadd_ps(v0_12, l24_12, acc0_24);
-__m256i ws0_13 = _mm256_srli_epi32(w0, 26);
-__m256i ws8_13 = _mm256_srli_epi32(w8, 26);
-__m256i ws16_13 = _mm256_srli_epi32(w16, 26);
-__m256i ws24_13 = _mm256_srli_epi32(w24, 26);
-__m256i wsa0_13= _mm256_and_si256(ws0_13, mask);
-__m256i wsa8_13= _mm256_and_si256(ws8_13, mask);
-__m256i wsa16_13= _mm256_and_si256(ws16_13, mask);
-__m256i wsa24_13= _mm256_and_si256(ws24_13, mask);
-__m256 l0_13 = _mm256_cvtepi32_ps(wsa0_13);
-__m256 l8_13 = _mm256_cvtepi32_ps(wsa8_13);
-__m256 l16_13 = _mm256_cvtepi32_ps(wsa16_13);
-__m256 l24_13 = _mm256_cvtepi32_ps(wsa24_13);
-acc0_0 = _mm256_fmadd_ps(v0_13, l0_13, acc0_0);
-acc0_8 = _mm256_fmadd_ps(v0_13, l8_13, acc0_8);
-acc0_16 = _mm256_fmadd_ps(v0_13, l16_13, acc0_16);
-acc0_24 = _mm256_fmadd_ps(v0_13, l24_13, acc0_24);
-__m256i ws0_14 = _mm256_srli_epi32(w0, 28);
-__m256i ws8_14 = _mm256_srli_epi32(w8, 28);
-__m256i ws16_14 = _mm256_srli_epi32(w16, 28);
-__m256i ws24_14 = _mm256_srli_epi32(w24, 28);
-__m256i wsa0_14= _mm256_and_si256(ws0_14, mask);
-__m256i wsa8_14= _mm256_and_si256(ws8_14, mask);
-__m256i wsa16_14= _mm256_and_si256(ws16_14, mask);
-__m256i wsa24_14= _mm256_and_si256(ws24_14, mask);
-__m256 l0_14 = _mm256_cvtepi32_ps(wsa0_14);
-__m256 l8_14 = _mm256_cvtepi32_ps(wsa8_14);
-__m256 l16_14 = _mm256_cvtepi32_ps(wsa16_14);
-__m256 l24_14 = _mm256_cvtepi32_ps(wsa24_14);
-acc0_0 = _mm256_fmadd_ps(v0_14, l0_14, acc0_0);
-acc0_8 = _mm256_fmadd_ps(v0_14, l8_14, acc0_8);
-acc0_16 = _mm256_fmadd_ps(v0_14, l16_14, acc0_16);
-acc0_24 = _mm256_fmadd_ps(v0_14, l24_14, acc0_24);
-__m256i ws0_15 = _mm256_srli_epi32(w0, 30);
-__m256i ws8_15 = _mm256_srli_epi32(w8, 30);
-__m256i ws16_15 = _mm256_srli_epi32(w16, 30);
-__m256i ws24_15 = _mm256_srli_epi32(w24, 30);
-__m256i wsa0_15= _mm256_and_si256(ws0_15, mask);
-__m256i wsa8_15= _mm256_and_si256(ws8_15, mask);
-__m256i wsa16_15= _mm256_and_si256(ws16_15, mask);
-__m256i wsa24_15= _mm256_and_si256(ws24_15, mask);
-__m256 l0_15 = _mm256_cvtepi32_ps(wsa0_15);
-__m256 l8_15 = _mm256_cvtepi32_ps(wsa8_15);
-__m256 l16_15 = _mm256_cvtepi32_ps(wsa16_15);
-__m256 l24_15 = _mm256_cvtepi32_ps(wsa24_15);
-acc0_0 = _mm256_fmadd_ps(v0_15, l0_15, acc0_0);
-acc0_8 = _mm256_fmadd_ps(v0_15, l8_15, acc0_8);
-acc0_16 = _mm256_fmadd_ps(v0_15, l16_15, acc0_16);
-acc0_24 = _mm256_fmadd_ps(v0_15, l24_15, acc0_24);
-__m256 v0_7 = _mm256_set1_ps(input[(i*om+k)*mb*nb + (k2+7)*nb + i1+0]);
-__m256 v0_6 = _mm256_set1_ps(input[(i*om+k)*mb*nb + (k2+6)*nb + i1+0]);
-__m256 v0_5 = _mm256_set1_ps(input[(i*om+k)*mb*nb + (k2+5)*nb + i1+0]);
-__m256 v0_4 = _mm256_set1_ps(input[(i*om+k)*mb*nb + (k2+4)*nb + i1+0]);
-__m256 v0_3 = _mm256_set1_ps(input[(i*om+k)*mb*nb + (k2+3)*nb + i1+0]);
-__m256 v0_2 = _mm256_set1_ps(input[(i*om+k)*mb*nb + (k2+2)*nb + i1+0]);
-__m256 v0_1 = _mm256_set1_ps(input[(i*om+k)*mb*nb + (k2+1)*nb + i1+0]);
-__m256 v0_0 = _mm256_set1_ps(input[(i*om+k)*mb*nb + (k2+0)*nb + i1+0]);
-__m256i ws0_0 = _mm256_srli_epi32(w0, 0);
-__m256i ws8_0 = _mm256_srli_epi32(w8, 0);
-__m256i ws16_0 = _mm256_srli_epi32(w16, 0);
-__m256i ws24_0 = _mm256_srli_epi32(w24, 0);
-__m256i wsa0_0= _mm256_and_si256(ws0_0, mask);
-__m256i wsa8_0= _mm256_and_si256(ws8_0, mask);
-__m256i wsa16_0= _mm256_and_si256(ws16_0, mask);
-__m256i wsa24_0= _mm256_and_si256(ws24_0, mask);
-__m256 l0_0 = _mm256_cvtepi32_ps(wsa0_0);
-__m256 l8_0 = _mm256_cvtepi32_ps(wsa8_0);
-__m256 l16_0 = _mm256_cvtepi32_ps(wsa16_0);
-__m256 l24_0 = _mm256_cvtepi32_ps(wsa24_0);
-acc0_0 = _mm256_fmadd_ps(v0_0, l0_0, acc0_0);
-acc0_8 = _mm256_fmadd_ps(v0_0, l8_0, acc0_8);
-acc0_16 = _mm256_fmadd_ps(v0_0, l16_0, acc0_16);
-acc0_24 = _mm256_fmadd_ps(v0_0, l24_0, acc0_24);
-__m256i ws0_1 = _mm256_srli_epi32(w0, 2);
-__m256i ws8_1 = _mm256_srli_epi32(w8, 2);
-__m256i ws16_1 = _mm256_srli_epi32(w16, 2);
-__m256i ws24_1 = _mm256_srli_epi32(w24, 2);
-__m256i wsa0_1= _mm256_and_si256(ws0_1, mask);
-__m256i wsa8_1= _mm256_and_si256(ws8_1, mask);
-__m256i wsa16_1= _mm256_and_si256(ws16_1, mask);
-__m256i wsa24_1= _mm256_and_si256(ws24_1, mask);
-__m256 l0_1 = _mm256_cvtepi32_ps(wsa0_1);
-__m256 l8_1 = _mm256_cvtepi32_ps(wsa8_1);
-__m256 l16_1 = _mm256_cvtepi32_ps(wsa16_1);
-__m256 l24_1 = _mm256_cvtepi32_ps(wsa24_1);
-acc0_0 = _mm256_fmadd_ps(v0_1, l0_1, acc0_0);
-acc0_8 = _mm256_fmadd_ps(v0_1, l8_1, acc0_8);
-acc0_16 = _mm256_fmadd_ps(v0_1, l16_1, acc0_16);
-acc0_24 = _mm256_fmadd_ps(v0_1, l24_1, acc0_24);
-__m256i ws0_2 = _mm256_srli_epi32(w0, 4);
-__m256i ws8_2 = _mm256_srli_epi32(w8, 4);
-__m256i ws16_2 = _mm256_srli_epi32(w16, 4);
-__m256i ws24_2 = _mm256_srli_epi32(w24, 4);
-__m256i wsa0_2= _mm256_and_si256(ws0_2, mask);
-__m256i wsa8_2= _mm256_and_si256(ws8_2, mask);
-__m256i wsa16_2= _mm256_and_si256(ws16_2, mask);
-__m256i wsa24_2= _mm256_and_si256(ws24_2, mask);
-__m256 l0_2 = _mm256_cvtepi32_ps(wsa0_2);
-__m256 l8_2 = _mm256_cvtepi32_ps(wsa8_2);
-__m256 l16_2 = _mm256_cvtepi32_ps(wsa16_2);
-__m256 l24_2 = _mm256_cvtepi32_ps(wsa24_2);
-acc0_0 = _mm256_fmadd_ps(v0_2, l0_2, acc0_0);
-acc0_8 = _mm256_fmadd_ps(v0_2, l8_2, acc0_8);
-acc0_16 = _mm256_fmadd_ps(v0_2, l16_2, acc0_16);
-acc0_24 = _mm256_fmadd_ps(v0_2, l24_2, acc0_24);
-__m256i ws0_3 = _mm256_srli_epi32(w0, 6);
-__m256i ws8_3 = _mm256_srli_epi32(w8, 6);
-__m256i ws16_3 = _mm256_srli_epi32(w16, 6);
-__m256i ws24_3 = _mm256_srli_epi32(w24, 6);
-__m256i wsa0_3= _mm256_and_si256(ws0_3, mask);
-__m256i wsa8_3= _mm256_and_si256(ws8_3, mask);
-__m256i wsa16_3= _mm256_and_si256(ws16_3, mask);
-__m256i wsa24_3= _mm256_and_si256(ws24_3, mask);
-__m256 l0_3 = _mm256_cvtepi32_ps(wsa0_3);
-__m256 l8_3 = _mm256_cvtepi32_ps(wsa8_3);
-__m256 l16_3 = _mm256_cvtepi32_ps(wsa16_3);
-__m256 l24_3 = _mm256_cvtepi32_ps(wsa24_3);
-acc0_0 = _mm256_fmadd_ps(v0_3, l0_3, acc0_0);
-acc0_8 = _mm256_fmadd_ps(v0_3, l8_3, acc0_8);
-acc0_16 = _mm256_fmadd_ps(v0_3, l16_3, acc0_16);
-acc0_24 = _mm256_fmadd_ps(v0_3, l24_3, acc0_24);
-__m256i ws0_4 = _mm256_srli_epi32(w0, 8);
-__m256i ws8_4 = _mm256_srli_epi32(w8, 8);
-__m256i ws16_4 = _mm256_srli_epi32(w16, 8);
-__m256i ws24_4 = _mm256_srli_epi32(w24, 8);
-__m256i wsa0_4= _mm256_and_si256(ws0_4, mask);
-__m256i wsa8_4= _mm256_and_si256(ws8_4, mask);
-__m256i wsa16_4= _mm256_and_si256(ws16_4, mask);
-__m256i wsa24_4= _mm256_and_si256(ws24_4, mask);
-__m256 l0_4 = _mm256_cvtepi32_ps(wsa0_4);
-__m256 l8_4 = _mm256_cvtepi32_ps(wsa8_4);
-__m256 l16_4 = _mm256_cvtepi32_ps(wsa16_4);
-__m256 l24_4 = _mm256_cvtepi32_ps(wsa24_4);
-acc0_0 = _mm256_fmadd_ps(v0_4, l0_4, acc0_0);
-acc0_8 = _mm256_fmadd_ps(v0_4, l8_4, acc0_8);
-acc0_16 = _mm256_fmadd_ps(v0_4, l16_4, acc0_16);
-acc0_24 = _mm256_fmadd_ps(v0_4, l24_4, acc0_24);
-__m256i ws0_5 = _mm256_srli_epi32(w0, 10);
-__m256i ws8_5 = _mm256_srli_epi32(w8, 10);
-__m256i ws16_5 = _mm256_srli_epi32(w16, 10);
-__m256i ws24_5 = _mm256_srli_epi32(w24, 10);
-__m256i wsa0_5= _mm256_and_si256(ws0_5, mask);
-__m256i wsa8_5= _mm256_and_si256(ws8_5, mask);
-__m256i wsa16_5= _mm256_and_si256(ws16_5, mask);
-__m256i wsa24_5= _mm256_and_si256(ws24_5, mask);
-__m256 l0_5 = _mm256_cvtepi32_ps(wsa0_5);
-__m256 l8_5 = _mm256_cvtepi32_ps(wsa8_5);
-__m256 l16_5 = _mm256_cvtepi32_ps(wsa16_5);
-__m256 l24_5 = _mm256_cvtepi32_ps(wsa24_5);
-acc0_0 = _mm256_fmadd_ps(v0_5, l0_5, acc0_0);
-acc0_8 = _mm256_fmadd_ps(v0_5, l8_5, acc0_8);
-acc0_16 = _mm256_fmadd_ps(v0_5, l16_5, acc0_16);
-acc0_24 = _mm256_fmadd_ps(v0_5, l24_5, acc0_24);
-__m256i ws0_6 = _mm256_srli_epi32(w0, 12);
-__m256i ws8_6 = _mm256_srli_epi32(w8, 12);
-__m256i ws16_6 = _mm256_srli_epi32(w16, 12);
-__m256i ws24_6 = _mm256_srli_epi32(w24, 12);
-__m256i wsa0_6= _mm256_and_si256(ws0_6, mask);
-__m256i wsa8_6= _mm256_and_si256(ws8_6, mask);
-__m256i wsa16_6= _mm256_and_si256(ws16_6, mask);
-__m256i wsa24_6= _mm256_and_si256(ws24_6, mask);
-__m256 l0_6 = _mm256_cvtepi32_ps(wsa0_6);
-__m256 l8_6 = _mm256_cvtepi32_ps(wsa8_6);
-__m256 l16_6 = _mm256_cvtepi32_ps(wsa16_6);
-__m256 l24_6 = _mm256_cvtepi32_ps(wsa24_6);
-acc0_0 = _mm256_fmadd_ps(v0_6, l0_6, acc0_0);
-acc0_8 = _mm256_fmadd_ps(v0_6, l8_6, acc0_8);
-acc0_16 = _mm256_fmadd_ps(v0_6, l16_6, acc0_16);
-acc0_24 = _mm256_fmadd_ps(v0_6, l24_6, acc0_24);
-__m256i ws0_7 = _mm256_srli_epi32(w0, 14);
-__m256i ws8_7 = _mm256_srli_epi32(w8, 14);
-__m256i ws16_7 = _mm256_srli_epi32(w16, 14);
-__m256i ws24_7 = _mm256_srli_epi32(w24, 14);
-__m256i wsa0_7= _mm256_and_si256(ws0_7, mask);
-__m256i wsa8_7= _mm256_and_si256(ws8_7, mask);
-__m256i wsa16_7= _mm256_and_si256(ws16_7, mask);
-__m256i wsa24_7= _mm256_and_si256(ws24_7, mask);
-__m256 l0_7 = _mm256_cvtepi32_ps(wsa0_7);
-__m256 l8_7 = _mm256_cvtepi32_ps(wsa8_7);
-__m256 l16_7 = _mm256_cvtepi32_ps(wsa16_7);
-__m256 l24_7 = _mm256_cvtepi32_ps(wsa24_7);
-acc0_0 = _mm256_fmadd_ps(v0_7, l0_7, acc0_0);
-acc0_8 = _mm256_fmadd_ps(v0_7, l8_7, acc0_8);
-acc0_16 = _mm256_fmadd_ps(v0_7, l16_7, acc0_16);
-acc0_24 = _mm256_fmadd_ps(v0_7, l24_7, acc0_24);
-}
-__m256 o0_0 = _mm256_loadu_ps(&output[base_output + j + (i1+0)*t + j1+0]);
-__m256 o0_8 = _mm256_loadu_ps(&output[base_output + j + (i1+0)*t + j1+8]);
-__m256 o0_16 = _mm256_loadu_ps(&output[base_output + j + (i1+0)*t + j1+16]);
-__m256 o0_24 = _mm256_loadu_ps(&output[base_output + j + (i1+0)*t + j1+24]);
-__m256 s0_0 = _mm256_loadu_ps(&scales[(k*mb+k1)/gs * t + base_output + j + j1+0]);
-__m256 s0_8 = _mm256_loadu_ps(&scales[(k*mb+k1)/gs * t + base_output + j + j1+8]);
-__m256 s0_16 = _mm256_loadu_ps(&scales[(k*mb+k1)/gs * t + base_output + j + j1+16]);
-__m256 s0_24 = _mm256_loadu_ps(&scales[(k*mb+k1)/gs * t + base_output + j + j1+24]);
-__m256 f0_0 = _mm256_fmadd_ps(acc0_0, s0_0, o0_0);
-__m256 f0_8 = _mm256_fmadd_ps(acc0_8, s0_8, o0_8);
-__m256 f0_16 = _mm256_fmadd_ps(acc0_16, s0_16, o0_16);
-__m256 f0_24 = _mm256_fmadd_ps(acc0_24, s0_24, o0_24);
-_mm256_storeu_ps(&output[base_output + j + (i1+0)*t + j1+0], f0_0);
-_mm256_storeu_ps(&output[base_output + j + (i1+0)*t + j1+8], f0_8);
-_mm256_storeu_ps(&output[base_output + j + (i1+0)*t + j1+16], f0_16);
-_mm256_storeu_ps(&output[base_output + j + (i1+0)*t + j1+24], f0_24);
-}
-}
-}
-}
-}
-}
-#pragma omp barrier
-const int ngs = m/gs;
-for (int i = 0; i < n; i++) {
-for (int j = 0; j < tt; j+=32){
-__m256 acc0 = _mm256_setzero_ps();
-__m256 acc8 = _mm256_setzero_ps();
-__m256 acc16 = _mm256_setzero_ps();
-__m256 acc24 = _mm256_setzero_ps();
-for (int i1 = 0; i1 < ngs; i1++){
-__m256 r = _mm256_set1_ps(sums[i*ngs + i1]);
-__m256 z0 = _mm256_loadu_ps(&zeros[base_output + i1* t + j + 0]);
-__m256 z8 = _mm256_loadu_ps(&zeros[base_output + i1* t + j + 8]);
-__m256 z16 = _mm256_loadu_ps(&zeros[base_output + i1* t + j + 16]);
-__m256 z24 = _mm256_loadu_ps(&zeros[base_output + i1* t + j + 24]);
-__m256 s0 = _mm256_loadu_ps(&scales[base_output + i1 * t + j + 0]);
-__m256 s8 = _mm256_loadu_ps(&scales[base_output + i1 * t + j + 8]);
-__m256 s16 = _mm256_loadu_ps(&scales[base_output + i1 * t + j + 16]);
-__m256 s24 = _mm256_loadu_ps(&scales[base_output + i1 * t + j + 24]);
-__m256 zs0 = _mm256_mul_ps(z0, s0);
-__m256 zs8 = _mm256_mul_ps(z8, s8);
-__m256 zs16 = _mm256_mul_ps(z16, s16);
-__m256 zs24 = _mm256_mul_ps(z24, s24);
-acc0 = _mm256_fmadd_ps(zs0, r, acc0);
-acc8 = _mm256_fmadd_ps(zs8, r, acc8);
-acc16 = _mm256_fmadd_ps(zs16, r, acc16);
-acc24 = _mm256_fmadd_ps(zs24, r, acc24);
-}
-__m256 o0 = _mm256_loadu_ps(&output[i*t + base_output + j + 0]);
-__m256 o8 = _mm256_loadu_ps(&output[i*t + base_output + j + 8]);
-__m256 o16 = _mm256_loadu_ps(&output[i*t + base_output + j + 16]);
-__m256 o24 = _mm256_loadu_ps(&output[i*t + base_output + j + 24]);
-__m256 b0 = _mm256_loadu_ps(&bias[base_output + j + 0]);
-__m256 b8 = _mm256_loadu_ps(&bias[base_output + j + 8]);
-__m256 b16 = _mm256_loadu_ps(&bias[base_output + j + 16]);
-__m256 b24 = _mm256_loadu_ps(&bias[base_output + j + 24]);
-__m256 o10 = _mm256_add_ps(o0, acc0);
-__m256 o18 = _mm256_add_ps(o8, acc8);
-__m256 o116 = _mm256_add_ps(o16, acc16);
-__m256 o124 = _mm256_add_ps(o24, acc24);
-__m256 o20 = _mm256_add_ps(o10, b0);
-__m256 o28 = _mm256_add_ps(o18, b8);
-__m256 o216 = _mm256_add_ps(o116, b16);
-__m256 o224 = _mm256_add_ps(o124, b24);
-_mm256_storeu_ps(&output[i*t + base_output + j + 0], o20);
-_mm256_storeu_ps(&output[i*t + base_output + j + 8], o28);
-_mm256_storeu_ps(&output[i*t + base_output + j + 16], o216);
-_mm256_storeu_ps(&output[i*t + base_output + j + 24], o224);
-}
-}
-}
-}
-inline void qforward(const float* __restrict__ input, 
- const int* __restrict__ W, 
-const float* __restrict__ scales, 
-const float* __restrict__ zeros, 
-const float* __restrict__ bias, 
-const float* __restrict__ sums, 
-float* __restrict__ output, 
-int n, 
- int m, 
- int t) {
-q2gemm_gs(input, W, scales, zeros, bias, sums, output, n, m, t, 1, 1024, 32, 512, 64, 9);
-}
-inline void pack_input(float* A, float* B){
-  // copy the full matrix A in blocked format into B
-  uint64_t idx = 0;
-  const int N = 1;
-  const int M = 4096;
-  const int nb = 1;
-  const int mb = 1024;
-  for(int i = 0; i < N; i+=nb){ 
-             for(int j = 0; j < M; j+=mb){
-                 for(int jj = j; jj < mymin(j+mb, M); jj++){
-                     for(int ii = i; ii < mymin(i+nb, N); ii++){
-                         B[idx] = A[ii*M+jj];
-                         idx++;
-                     }
-                 }
-             }
-         }
-     }
-inline void pack_qw_inner(int* A, int* B, int cutoff){
-  // copy the full matrix A in blocked format into B
-  uint64_t idx = 0;
-  const int N = 256;
-  const int M = 4096;
-  const int nb = 64;
-int mb = 32;
-    for(int j = 0, tid = 0; j < M; j+=mb, tid++){
- for(int i = 0; i < N; i+=nb){
-                     for(int ii = i; ii < mymin(i+nb, N); ii++){
-                         for(int jj = j; jj < mymin(j+mb, M); jj++){
-                             B[idx] = A[ii*M+jj];
-                             idx++;
-                         }
-                     }
-                 }
-}
-}
-inline void pack_qw(int* A, int* B){
-  pack_qw_inner(A, B, 65);
-}
-inline void pack_output(float* A, float* B){
-  // copy the full matrix A in blocked format into B
-  uint64_t idx = 0;
-  const int N = 1;
-  const int M = 4096;
-  const int nb = 1;
-  const int mb = 32;
-  for(int i = 0; i < N; i+=nb){ 
-             for(int j = 0; j < M; j+=mb){
-                 for(int ii = i; ii < mymin(i+nb, N); ii++){
-                     for(int jj = j; jj < mymin(j+mb, M); jj++){
-                         B[idx] = A[ii*M+jj];
-                         idx++;
-                     }
-                 }
-             }
-         }
-     }
-void print_parameters(){
-std::ofstream outfile;
-outfile.open("./autogptq_extension/qigen/tmp.csv", std::ios_base::app);
-outfile << 2 << "," << 1 << "," << 16 << "," << 32 << "," << 8 << "," << 8  << "," << 64 << ",";
-}

autogptq_extension/qigen/intrin.py

@@ -1,149 +0,0 @@
-
-def load_int(to, address, const=True):
-    if const:
-        return f"const __m256i {to} = _mm256_loadu_si256({address});"
-    else:
-        return f"__m256i {to} = _mm256_loadu_si256({address});"
-
-def load_fp(to, address, const=True):
-    if const:
-        return f"const __m256 {to} = _mm256_loadu_ps({address});"
-    else:
-        return f"__m256 {to} = _mm256_loadu_ps({address});"
-
-# to = a * b + c
-def vfma(to, a, b, c):
-    return f"__m256 {to} = _mm256_fmadd_ps({a}, {b}, {c});"
-
-def vsrli(to, a, b):
-    return f"const __m256i {to} = _mm256_srli_epi32({a}, {b});"
-
-def vand(to, a, b):
-    return f"const __m256i {to} = _mm256_and_si256({a}, {b});"
-
-def vbroadcast_fp(to, a):
-    return f"const __m256 {to} = _mm256_set1_ps({a});"
-
-def vbroadcast_int32(to, a):
-    return f"__m256i {to} = _mm256_set1_epi32({a});"
-
-def vsetzero(to):
-    return f"__m256 {to} = _mm256_setzero_ps();"
-
-def vcvtepi32_ps(to, a):
-    return f"const __m256 {to} = _mm256_cvtepi32_ps({a});"
-
-def _256extractf128_ps(to, a, imm):
-    return f"const __m128 {to} = _mm256_extractf128_ps({a}, {imm});"
-
-def _256castps256_ps128(to, a):
-    return f"const __m128 {to} = _mm256_castps256_ps128({a});"
-
-def _add_ps(to, a, b):
-    return f"const __m128 {to} = _mm_add_ps({a}, {b});"
-
-def _movehl_ps(to, a, b):
-    return f"const __m128 {to} = _mm_movehl_ps({a}, {b});"
-
-def _shuffle_ps(to, a, b, imm):
-    return f"const __m128 {to} = _mm_shuffle_ps({a}, {b}, {imm});"
-
-def _cvtss_f32(to, a):
-    return f"const float {to} = _mm_cvtss_f32({a});"
-
-def _reduce8_acc(a, b, c, d, e, f, g, h):
-    res = ""
-    res += _256extractf128_ps("hi_quad0", a, 1)
-    res += _256extractf128_ps("hi_quad1", b, 1)
-    res += _256extractf128_ps("hi_quad2", c, 1)
-    res += _256extractf128_ps("hi_quad3", d, 1)
-    res += _256extractf128_ps("hi_quad4", e, 1)
-    res += _256extractf128_ps("hi_quad5", f, 1)
-    res += _256extractf128_ps("hi_quad6", g, 1)
-    res += _256extractf128_ps("hi_quad7", h, 1)
-
-    res += _256castps256_ps128("lo_quad0", a)
-    res += _256castps256_ps128("lo_quad1", b)
-    res += _256castps256_ps128("lo_quad2", c)
-    res += _256castps256_ps128("lo_quad3", d)
-    res += _256castps256_ps128("lo_quad4", e)
-    res += _256castps256_ps128("lo_quad5", f)
-    res += _256castps256_ps128("lo_quad6", g)
-    res += _256castps256_ps128("lo_quad7", h)
-
-    res += _add_ps("sum_quad0", "lo_quad0", "hi_quad0")
-    res += _add_ps("sum_quad1", "lo_quad1", "hi_quad1")
-    res += _add_ps("sum_quad2", "lo_quad2", "hi_quad2")
-    res += _add_ps("sum_quad3", "lo_quad3", "hi_quad3")
-    res += _add_ps("sum_quad4", "lo_quad4", "hi_quad4")
-    res += _add_ps("sum_quad5", "lo_quad5", "hi_quad5")
-    res += _add_ps("sum_quad6", "lo_quad6", "hi_quad6")
-    res += _add_ps("sum_quad7", "lo_quad7", "hi_quad7")
-
-    res += _movehl_ps("hi_dual0", "sum_quad0", "sum_quad0")
-    res += _movehl_ps("hi_dual1", "sum_quad1", "sum_quad1")
-    res += _movehl_ps("hi_dual2", "sum_quad2", "sum_quad2")
-    res += _movehl_ps("hi_dual3", "sum_quad3", "sum_quad3")
-    res += _movehl_ps("hi_dual4", "sum_quad4", "sum_quad4")
-    res += _movehl_ps("hi_dual5", "sum_quad5", "sum_quad5")
-    res += _movehl_ps("hi_dual6", "sum_quad6", "sum_quad6")
-    res += _movehl_ps("hi_dual7", "sum_quad7", "sum_quad7")
-
-    res += _add_ps("sum_dual0", "sum_quad0", "hi_dual0")
-    res += _add_ps("sum_dual1", "sum_quad1", "hi_dual1")
-    res += _add_ps("sum_dual2", "sum_quad2", "hi_dual2")
-    res += _add_ps("sum_dual3", "sum_quad3", "hi_dual3")
-    res += _add_ps("sum_dual4", "sum_quad4", "hi_dual4")
-    res += _add_ps("sum_dual5", "sum_quad5", "hi_dual5")
-    res += _add_ps("sum_dual6", "sum_quad6", "hi_dual6")
-    res += _add_ps("sum_dual7", "sum_quad7", "hi_dual7")
-
-    res += _shuffle_ps("hi0", "sum_dual0", "sum_dual0", 0x1)
-    res += _shuffle_ps("hi1", "sum_dual1", "sum_dual1", 0x1)
-    res += _shuffle_ps("hi2", "sum_dual2", "sum_dual2", 0x1)
-    res += _shuffle_ps("hi3", "sum_dual3", "sum_dual3", 0x1)
-    res += _shuffle_ps("hi4", "sum_dual4", "sum_dual4", 0x1)
-    res += _shuffle_ps("hi5", "sum_dual5", "sum_dual5", 0x1)
-    res += _shuffle_ps("hi6", "sum_dual6", "sum_dual6", 0x1)
-    res += _shuffle_ps("hi7", "sum_dual7", "sum_dual7", 0x1)
-    
-    res += _add_ps("sum0", "sum_dual0", "hi0")
-    res += _add_ps("sum1", "sum_dual1", "hi1")
-    res += _add_ps("sum2", "sum_dual2", "hi2")
-    res += _add_ps("sum3", "sum_dual3", "hi3")
-    res += _add_ps("sum4", "sum_dual4", "hi4")
-    res += _add_ps("sum5", "sum_dual5", "hi5")
-    res += _add_ps("sum6", "sum_dual6", "hi6")
-    res += _add_ps("sum7", "sum_dual7", "hi7")
-
-    res += _cvtss_f32(f"f{a}", "sum0")
-    res += _cvtss_f32(f"f{b}", "sum1")
-    res += _cvtss_f32(f"f{c}", "sum2")
-    res += _cvtss_f32(f"f{d}", "sum3")
-    res += _cvtss_f32(f"f{e}", "sum4")
-    res += _cvtss_f32(f"f{f}", "sum5")
-    res += _cvtss_f32(f"f{g}", "sum6")
-    res += _cvtss_f32(f"f{h}", "sum7")
-
-    return res
-
-acc_idx = 0
-def _reduce_add(a):
-    global acc_idx
-    res = ""
-    res += _256extractf128_ps(f"hi_quad{acc_idx}", a, 1)
-    res += _256castps256_ps128(f"lo_quad{acc_idx}", a)
-    res += _add_ps(f"sum_quad{acc_idx}", f"lo_quad{acc_idx}", f"hi_quad{acc_idx}")
-    res += _movehl_ps(f"hi_dual{acc_idx}", f"sum_quad{acc_idx}", f"sum_quad{acc_idx}")
-    res += _add_ps(f"sum_dual{acc_idx}", f"sum_quad{acc_idx}", f"hi_dual{acc_idx}")
-    res += _shuffle_ps(f"hi{acc_idx}", f"sum_dual{acc_idx}", f"sum_dual{acc_idx}", 0x1)
-    res += _add_ps(f"sum{acc_idx}", f"sum_dual{acc_idx}", f"hi{acc_idx}")
-    res += _cvtss_f32(f"f{a}", f"sum{acc_idx}")
-    acc_idx += 1
-    return res
-
-
-
-
-
-

autogptq_extension/qigen/mmm.cpp

@@ -1,302 +0,0 @@
-#include <iostream>
-#include "forward.h"
-#include <cstring>
-#include <algorithm>
-#include <vector>
-#include <chrono>
-#include <fstream>
-
-#define mymin(a,b) ((a)<(b)?(a):(b))
-#define mymax(a,b) ((a)>(b)?(a):(b))
-
-void print_matrix(std::string name, float* A, int N, int M){
-	std::cout<<name<<std::endl;
-	for(int i = 0; i < N; i++){
-		for(int j = 0; j < M; j++){
-			std::cout << A[i*M+j] << " ";
-		}
-		std::cout << std::endl;
-	}
-	std::cout<<std::endl;
-}
-
-void oracle_mmadd(float* A, float* B, float* bias, float* C, int n, int m, int t){
-	// triple loop matmul and add bias
-	for (int i = 0; i < n; i++){
-		for (int j = 0; j < t; j++){
-			float sum = 0;
-			for (int k = 0; k < m; k++){
-				sum += A[i*m+k] * B[k*t+j];
-			}
-			C[i*t+j] += sum + bias[j];
-		}
-	}
-}
-
-void compute_reduction(float *in, float *out, int n, int m, int gs){
-	int ng;
-	if(gs == -1){
-		ng = 1;
-		gs = m;
-	}else{
-		ng = m/gs;
-	}
-	for(int i = 0; i < n; i++){
-		for(int j0 = 0; j0 < m; j0+=gs){
-			int j = j0/gs;
-			out[i*ng+j] = 0;
-			for(int j1 = j0; j1 < j0+gs; j1++){
-				out[i*ng+j] += in[i*m+j1];
-			}
-		}
-	}
-}
-
-void quantize_sim(float* A, float* BQ, float* scales, float* zeros, int n, int m, int bits, int gs){
-	//find scales and zeros arrays
-	if(gs == -1){
-		gs = n;
-	}
-	float range = (1<<bits) - 1;
-	int packed = 32 / bits;
-
-	for(int i0 = 0; i0 < n; i0+=gs){
-		int row = i0/gs;
-		for(int j = 0; j < m; j++){
-			float min = A[i0*m + j];
-			float max = A[i0*m + j];
-			for(int i1 = i0; i1 < i0+gs; i1++){
-				min = mymin(min, A[i1*m+j]);
-				max = mymax(max, A[i1*m+j]);
-			}
-			scales[row*m + j] = (max-min)/range;
-			zeros[row*m + j ] = min;
-		}
-		for(int j = 0; j < m; j++){
-			for (int i1 = i0; i1 < i0+gs; i1++){
-				uint32_t acc = 0;
-				int temp = (A[i1*m+j] - zeros[row*m+j])/scales[row*m+j];
-				float val = ((float) temp + zeros[row*m+j]) * scales[row*m+j];
-				BQ[i1*m+j] = val;
-			}
-		}
-	}
-
-}
-
-void quantize(float* A, int* BQ, float* scales, float* zeros, int n, int m, int bits, int gs){
-	//find scales and zeros arrays
-	if(gs == -1){
-		gs = n;
-	}
-	float range = (1<<bits) - 1;
-	int packed = 32 / bits;
-
-	for(int i0 = 0; i0 < n; i0+=gs){
-		int row = i0/gs;
-		for(int j = 0; j < m; j++){
-			float min = A[i0*m + j];
-			float max = A[i0*m + j];
-			for(int i1 = i0; i1 < i0+gs; i1++){
-				min = mymin(min, A[i1*m+j]);
-				max = mymax(max, A[i1*m+j]);
-			}
-			scales[row*m + j] = (max-min)/range;
-			zeros[row*m + j ] = min;
-		}
-		for(int j = 0; j < m; j++){
-			if(bits == 3){
-				for (int i1 = i0; i1 < i0+gs; i1+=32){
-					uint32_t acc = 0;
-					int temp0 = ((int)((A[(i1+0)*m+j] - zeros[row*m+j])/scales[row*m+j])) << 0;
-					int temp1 = ((int)((A[(i1+1)*m+j] - zeros[row*m+j])/scales[row*m+j])) << 3;
-					int temp2 = ((int)((A[(i1+2)*m+j] - zeros[row*m+j])/scales[row*m+j])) << 6;
-					int temp3 = ((int)((A[(i1+3)*m+j] - zeros[row*m+j])/scales[row*m+j])) << 9;
-					int temp4 = ((int)((A[(i1+4)*m+j] - zeros[row*m+j])/scales[row*m+j])) << 12;
-					int temp5 = ((int)((A[(i1+5)*m+j] - zeros[row*m+j])/scales[row*m+j])) << 15;
-					int temp6 = ((int)((A[(i1+6)*m+j] - zeros[row*m+j])/scales[row*m+j])) << 18;
-					int temp7 = ((int)((A[(i1+7)*m+j] - zeros[row*m+j])/scales[row*m+j])) << 21;
-					int temp8 = ((int)((A[(i1+8)*m+j] - zeros[row*m+j])/scales[row*m+j])) << 24;
-					int temp9 = ((int)((A[(i1+9)*m+j] - zeros[row*m+j])/scales[row*m+j])) << 27;
-					int temp10_0 = ((int)((A[(i1+10)*m+j] - zeros[row*m+j])/scales[row*m+j])) << 30;
-					int temp10_1 = ((int)((A[(i1+10)*m+j] - zeros[row*m+j])/scales[row*m+j])) >> 2;
-					int temp11 = ((int)((A[(i1+11)*m+j] - zeros[row*m+j])/scales[row*m+j])) << 1;
-					int temp12 = ((int)((A[(i1+12)*m+j] - zeros[row*m+j])/scales[row*m+j])) << 4;
-					int temp13 = ((int)((A[(i1+13)*m+j] - zeros[row*m+j])/scales[row*m+j])) << 7;
-					int temp14 = ((int)((A[(i1+14)*m+j] - zeros[row*m+j])/scales[row*m+j])) << 10;
-					int temp15 = ((int)((A[(i1+15)*m+j] - zeros[row*m+j])/scales[row*m+j])) << 13;
-					int temp16 = ((int)((A[(i1+16)*m+j] - zeros[row*m+j])/scales[row*m+j])) << 16;
-					int temp17 = ((int)((A[(i1+17)*m+j] - zeros[row*m+j])/scales[row*m+j])) << 19;
-					int temp18 = ((int)((A[(i1+18)*m+j] - zeros[row*m+j])/scales[row*m+j])) << 22;
-					int temp19 = ((int)((A[(i1+19)*m+j] - zeros[row*m+j])/scales[row*m+j])) << 25;
-					int temp20 = ((int)((A[(i1+20)*m+j] - zeros[row*m+j])/scales[row*m+j])) << 28;
-					int temp21_0 = ((int)((A[(i1+21)*m+j] - zeros[row*m+j])/scales[row*m+j])) << 31;
-					int temp21_1 = ((int)((A[(i1+21)*m+j] - zeros[row*m+j])/scales[row*m+j])) >> 1;
-					int temp22 = ((int)((A[(i1+22)*m+j] - zeros[row*m+j])/scales[row*m+j])) << 2;
-					int temp23 = ((int)((A[(i1+23)*m+j] - zeros[row*m+j])/scales[row*m+j])) << 5;
-					int temp24 = ((int)((A[(i1+24)*m+j] - zeros[row*m+j])/scales[row*m+j])) << 8;
-					int temp25 = ((int)((A[(i1+25)*m+j] - zeros[row*m+j])/scales[row*m+j])) << 11;
-					int temp26 = ((int)((A[(i1+26)*m+j] - zeros[row*m+j])/scales[row*m+j])) << 14;
-					int temp27 = ((int)((A[(i1+27)*m+j] - zeros[row*m+j])/scales[row*m+j])) << 17;
-					int temp28 = ((int)((A[(i1+28)*m+j] - zeros[row*m+j])/scales[row*m+j])) << 20;
-					int temp29 = ((int)((A[(i1+29)*m+j] - zeros[row*m+j])/scales[row*m+j])) << 23;
-					int temp30 = ((int)((A[(i1+30)*m+j] - zeros[row*m+j])/scales[row*m+j])) << 26;
-					int temp31 = ((int)((A[(i1+31)*m+j] - zeros[row*m+j])/scales[row*m+j])) << 29;
-
-					int acc0 = 0, acc1 = 0, acc2 = 0;
-					
-					acc0 |= temp0;
-					acc0 |= temp1;
-					acc0 |= temp2;
-					acc0 |= temp3;
-					acc0 |= temp4;
-					acc0 |= temp5;
-					acc0 |= temp6;
-					acc0 |= temp7;
-					acc0 |= temp8;
-					acc0 |= temp9;
-					acc0 |= temp10_0;
-
-					acc1 |= temp10_1;
-					acc1 |= temp11;
-					acc1 |= temp12;
-					acc1 |= temp13;
-					acc1 |= temp14;
-					acc1 |= temp15;
-					acc1 |= temp16;
-					acc1 |= temp17;
-					acc1 |= temp18;
-					acc1 |= temp19;
-					acc1 |= temp20;
-					acc1 |= temp21_0;
-
-					acc2 |= temp21_1;
-					acc2 |= temp22;
-					acc2 |= temp23;
-					acc2 |= temp24;
-					acc2 |= temp25;
-					acc2 |= temp26;
-					acc2 |= temp27;
-					acc2 |= temp28;
-					acc2 |= temp29;
-					acc2 |= temp30;
-					acc2 |= temp31;
-
-					BQ[(3*i1/32)*m+j] = acc0;
-					BQ[(3*i1/32+1)*m+j] = acc1;
-					BQ[(3*i1/32+2)*m+j] = acc2;
-			}
-
-			}else{
-				for (int i1 = i0; i1 < i0+gs; i1+=packed){
-					uint32_t acc = 0;
-					for (int i2 = i1; i2 < i1+packed; i2++){
-						int temp = (A[i2*m+j] - zeros[row*m+j])/scales[row*m+j];
-						acc = acc | (temp << (bits*(i2-i1)));
-					}
-					BQ[(i1/packed)*m+j] = acc;
-				}
-			}
-		}
-	}
-
-}
-
-int main(int argc, char *argv[]){
-	// read n m t from args
-	if(argc == 0){std::cout << "Parameters not given\n"; return 0;}
-	int n = atoi(argv[1]);
-	int m = atoi(argv[2]);
-	int t = atoi(argv[3]);
-	int bits = atoi(argv[4]);
-	int gs = atoi(argv[5]);
-	int ng;
-	if(gs == -1){
-		ng = 1;
-	}else{
-		ng = m/gs;
-	}
-	float* A = new float[n*m];
-	float* AB = new float[n*m];
-	float* B = new float[m*t];
-	float* BQS = new float[m*t];
-	float* scales = new float[t*ng];
-	float* zeros = new float[t*ng];
-	int* BQ = new int[m*t/8];
-	int* BQB = new int[m*t/8];
-	float* sums = new float[n*ng];
-	float* bias = new float[t];
-	float* C = new float[n*t];
-	float* CB = new float[n*t];
-	float* C2 = new float[n*t];
-	srand(1);
-	for (int i = 0; i < n*m; i++){
-		A[i] = (float)rand() / RAND_MAX;
-	}
-	for (int i = 0; i < t*m; i++){
-		B[i] = (float)rand() / RAND_MAX;
-	}
-	for (int i = 0; i < t; i++){
-		bias[i] = (float)rand() / RAND_MAX;
-	}
-	for (int i = 0; i < n*t; i++){
-		C[i] = 0.0;
-		C2[i] = 0.0;
-	}
-	quantize_sim(B,BQS,scales,zeros,m,t,bits,gs);
-	quantize(B,BQ,scales,zeros,m,t,bits,gs);
-
-	quantize_sim(B,BQS,scales,zeros,m,t,bits,gs);
-	quantize(B,BQ,scales,zeros,m,t,bits,gs);
-	oracle_mmadd(A, BQS, bias, C, n, m, t);
-	pack_input(A,AB);
-	pack_qw(BQ,BQB);
-	pack_output(C,CB);
-
-	compute_reduction(A,sums,n,m,gs);
-	qforward(AB,BQB,scales,zeros,bias,sums,C2,n,m,t);
-
-	float norm = 0.0;
-	for (int i = 0; i < n*t; i++){
-		norm += (C[i] - C2[i]) * (C[i] - C2[i]); 
-	}
-	if(norm / (n*t) < 0.0001){
-		int iter = 30;
-		for(int _ = 0; _ < iter; _++){
-			qforward(AB,BQB,scales,zeros,bias,sums,C2,n,m,t);
-		}
-
-		int num_runs = 15;
-		std::vector<long int> runs(num_runs);
-		for(int r = 0; r < num_runs; r++){
-			auto start = std::chrono::high_resolution_clock::now();
-			for(int _ = 0; _ < iter; _++){
-				qforward(AB,BQB,scales,zeros,bias,sums,C2,n,m,t);
-			}
-			auto end = std::chrono::high_resolution_clock::now();
-			runs[r] = std::chrono::duration_cast<std::chrono::nanoseconds>(end - start).count();
-
-		}
-
-		std::sort(runs.begin(), runs.end());
-
-		float cycles_final = runs[num_runs/2 + 1] / iter;
-
-		std::ofstream outfile;
-		outfile.open("./autogptq_extension/qigen/tmp.csv", std::ios_base::app);
-
-		print_parameters();
-		outfile << cycles_final << std::endl;
-	}else{
-		float cycles_final = int(10e12);
-
-		std::ofstream outfile;
-		outfile.open("./autogptq_extension/qigen/tmp.csv", std::ios_base::app);
-
-		print_parameters();
-		outfile << cycles_final << std::endl;
-	}
-
-	return 0;	
-}
-

autogptq_extension/qigen/template.py

@@ -1,85 +0,0 @@
-
-def includes():
-    out = " \
-#include <torch/all.h>\n \
-#include <torch/python.h>\n \
-#include <omp.h>\n \
-#include <cmath>\n \
-#include <immintrin.h>\n \
-\n \
-#define mymin(a,b) ((a)<(b)?(a):(b))\n \
-#define mymax(a,b) ((a)>(b)?(a):(b))\n \
-"
-    return out
-
-
-def module(bits_list=[4, 2]):
-    out = 'PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {\n'
-    for bits in bits_list:
-        out += '  m.def("forward{}", &forward{}_cpu);\n'.format(bits, bits)
-
-    for bits in bits_list:
-        out += '  m.def("unpack_zeros{}", &unpack_zeros{});\n'.format(bits, bits)
-    
-    for bits in bits_list:
-        out += '  m.def("forward_gs{}", &forward{}_gs_cpu);\n'.format(bits, bits)
-    
-    for bits in bits_list:
-        out += '  m.def("pack{}", &pack{}_w_cpu);\n'.format(bits, bits)
-
-    out += 'm.def("compute_reduction_cpp", &compute_reduction);\n'
-    out += 'm.def("unquantize_sim", &unquantize_sim);\n'
-
-    # if oracle:
-        # out += '  m.def("forward4_oracle", &forward4_oracle_cpu);\n'
-
-
-    out += 'm.def("quant_scalar_scaled", &quant_scalar_cpu);\n'
-
-    out += '}\n'
-    return out
-
-def quant_scalar():
-    out = " \
-void quantize_scalar(float* A, int* BQ, float* scales, float* zeros, int n, int m, int bits){ \n \
-	//find scales and zeros arrays \n \
-	//quantize \n \
-	int pack = 32/bits;\n \
-	for (int j = 0; j < m; j++){\n \
-		for (int i = 0; i < n; i+=pack){\n \
-			uint32_t acc = 0;\n \
-			for (int ii = i; ii < i+pack; ii++){\n \
-				float ftemp = std::round((A[ii*m+j] + zeros[j])/scales[j]);\n \
-				int temp = (int)ftemp;\n \
-				acc = acc | (temp << (bits*(ii-i)));\n \
-			}\n \
-			BQ[(i/pack)*m+j] = acc;\n \
-			//BQ[0] = acc;\n \
-		}\n \
-	}\n \
-}\n \
-\n \
-void quant_scalar_cpu(\n \
-	torch::Tensor in, torch::Tensor out, \n \
-	torch::Tensor scales, torch::Tensor zeros, int bits\n \
-) {\n \
-\n \
-	int N  = in.size(0);\n \
-	int M  = in.size(1);\n \
-\n \
-	float* input = in.data_ptr<float>(); \n \
-	float* s   = scales.data_ptr<float>();\n \
-	float* z   = zeros.data_ptr<float>();\n \
-	int* O   = out.data_ptr<int>();\n \
-		\n \
-	quantize_scalar(input, O, s, z, N, M, bits);\n \
-\n \
-}\n"
-
-    return out
-
-
-
-
-
-

autogptq_extension/qigen/tmp.csv

@@ -1,37 +0,0 @@
-bits,nu,mu,tu,unroll,p,gs,time
-4,1,16,16,1,8,-1,1.3814e+06
-4,1,16,16,2,8,-1,1.44087e+06
-4,1,16,16,4,8,-1,1.56173e+06
-4,1,16,16,8,8,-1,1.41389e+06
-3,1,16,16,5,8,-1,2.14748e+09
-2,1,16,16,1,8,-1,1.09513e+06
-2,1,16,16,2,8,-1,1.11322e+06
-2,1,16,16,4,8,-1,1.12031e+06
-2,1,16,16,8,8,-1,1.19086e+06
-4,1,16,16,1,8,64,1.69111e+06
-4,1,16,16,2,8,64,1.60056e+06
-4,1,16,16,4,8,64,1.41263e+06
-4,1,16,16,8,8,64,1.74572e+06
-3,1,16,16,5,8,64,1.48062e+06
-2,1,16,16,1,8,64,1.51234e+06
-2,1,16,16,2,8,64,1.68108e+06
-2,1,16,16,4,8,64,1.7624e+06
-2,1,16,16,8,8,64,1.69563e+06
-4,1,16,32,1,8,-1,1.24798e+06
-4,1,16,32,2,8,-1,1.58421e+06
-4,1,16,32,4,8,-1,2.10718e+06
-4,1,16,32,8,8,-1,1.54288e+06
-3,1,16,32,5,8,-1,2.14748e+09
-2,1,16,32,1,8,-1,1.55906e+06
-2,1,16,32,2,8,-1,1.58576e+06
-2,1,16,32,4,8,-1,1.57993e+06
-2,1,16,32,8,8,-1,1.80443e+06
-4,1,16,32,1,8,64,1.58354e+06
-4,1,16,32,2,8,64,1.63248e+06
-4,1,16,32,4,8,64,1.91902e+06
-4,1,16,32,8,8,64,1.9243e+06
-3,1,16,32,5,8,64,1.33812e+06
-2,1,16,32,1,8,64,1.77522e+06
-2,1,16,32,2,8,64,1.54702e+06
-2,1,16,32,4,8,64,1.78772e+06
-2,1,16,32,8,8,64,1.49612e+06

docs/NEWS_OR_UPDATE.md

@@ -1,7 +1,5 @@
 ## <center>News or Update</center>
 
-- 2023-08-23 - (News) - 🤗 Transformers, optimum and peft have integrated `auto-gptq`, so now running and training GPTQ models can be more available to everyone! See [this blog](https://huggingface.co/blog/gptq-integration) and it's resources for more details!
-- 2023-08-21 - (News) - Team of Qwen officially released 4bit quantized version of Qwen-7B based on `auto-gptq`, and provided [a detailed benchmark results](https://huggingface.co/Qwen/Qwen-7B-Chat-Int4#%E9%87%8F%E5%8C%96-quantization)
 - 2023-08-06 - (Update) - Support exllama's q4 CUDA kernel to have at least 1.3x speed up for int4 quantized models when doing inference.
 - 2023-08-04 - (Update) - Support RoCm so that AMD GPU users can use auto-gptq with CUDA extensions.
 - 2023-07-26 - (Update) - An elegant [PPL benchmark script](examples/benchmark/perplexity.py) to get results that can be fairly compared with other libraries such as `llama.cpp`.

examples/benchmark/generation_speed.py

@@ -146,8 +146,7 @@ def load_model_tokenizer(
     use_safetensors: bool = False,
     use_fast_tokenizer: bool = False,
     inject_fused_attention: bool = True,
-    inject_fused_mlp: bool = True,
-    disable_exllama: bool = False
+    inject_fused_mlp: bool = True
 ):
     tokenizer = AutoTokenizer.from_pretrained(
         pretrained_model_name_or_path=tokenizer_name_or_path or model_name_or_path,
@@ -177,8 +176,7 @@ def load_model_tokenizer(
             model_basename=model_basename,
             use_safetensors=use_safetensors,
             trust_remote_code=trust_remote_code,
-            warmup_triton=False,
-            disable_exllama=disable_exllama
+            warmup_triton=False
         )
 
     return model, tokenizer
@@ -236,7 +234,6 @@ def main():
     parser.add_argument("--use_triton", action="store_true")
     parser.add_argument("--use_safetensors", action="store_true")
     parser.add_argument("--use_fast_tokenizer", action="store_true")
-    parser.add_argument("--disable_exllama", action="store_true")
     parser.add_argument("--no_inject_fused_attention", action="store_true")
     parser.add_argument("--no_inject_fused_mlp", action="store_true")
     parser.add_argument("--num_samples", type=int, default=10)
@@ -278,8 +275,7 @@ def main():
         use_safetensors=args.use_safetensors,
         use_fast_tokenizer=args.use_fast_tokenizer,
         inject_fused_attention=not args.no_inject_fused_attention,
-        inject_fused_mlp=not args.no_inject_fused_mlp,
-        disable_exllama=args.disable_exllama
+        inject_fused_mlp=not args.no_inject_fused_mlp
     )
     end = time.time()
     logger.info(f"model and tokenizer loading time: {end - start:.4f}s")

examples/benchmark/perplexity.py

@@ -37,7 +37,6 @@ if __name__ == "__main__":
     parser.add_argument("--use_safetensors", action="store_true", help="Whether to use safetensors model file")
     parser.add_argument("--use_fast_tokenizer", action="store_true", help="Wheter to use fast tokenizer")
     parser.add_argument("--trust_remote_code", action="store_true", help="Whether to use remote code")
-    parser.add_argument("--disable_exllama", action="store_true", help="Whether to use disable exllama kernel")
     args = parser.parse_args()
 
     os.environ["TOKENIZERS_PARALLELISM"] = "false"
@@ -69,8 +68,7 @@ if __name__ == "__main__":
             use_safetensors=args.use_safetensors,
             trust_remote_code=args.trust_remote_code,
             inject_fused_mlp=False,
-            inject_fused_attention=False,
-            disable_exllama=args.disable_exllama
+            inject_fused_attention=False
         )
     else:
         from transformers import AutoModelForCausalLM

setup.py

@@ -1,93 +1,78 @@
 import os
+import platform
 import sys
 from pathlib import Path
-from setuptools import setup, Extension, find_packages
-import subprocess
-import math
-import platform
+from setuptools import setup, find_packages
 
-os.environ["CC"] = "g++"
-os.environ["CXX"] = "g++"
+python_min_version = (3, 8, 0)
+python_min_version_str = '.'.join(map(str, python_min_version))
+if sys.version_info < python_min_version:
+    print(f"You are using Python {platform.python_version()}. Python >={python_min_version_str} is required.")
+    sys.exit(-1)
+
+BUILD_CUDA_EXT = int(os.environ.get('BUILD_CUDA_EXT', '1')) == 1
+
+if BUILD_CUDA_EXT:
+    try:
+        import torch
+    except:
+        print("torch is not installed, please install torch first!")
+        sys.exit(-1)
+
+    CUDA_VERSION = False
+    ROCM_VERSION = os.environ.get('ROCM_VERSION', False)
+    if ROCM_VERSION and not torch.version.hip:
+        raise ValueError(f"Trying to compile AutoGPTQ for RoCm, but PyTorch {torch.__version__} is installed with no RoCm support.")
+
+    if not ROCM_VERSION:
+        default_cuda_version = "".join(torch.version.cuda.split("."))
+        CUDA_VERSION = os.environ.get("CUDA_VERSION", default_cuda_version)
 
 common_setup_kwargs = {
-    "version": "0.5.0.dev0",
+    "version": "0.3.2",
     "name": "auto_gptq",
     "author": "PanQiWei",
     "description": "An easy-to-use LLMs quantization package with user-friendly apis, based on GPTQ algorithm.",
     "long_description": (Path(__file__).parent / "README.md").read_text(encoding="UTF-8"),
     "long_description_content_type": "text/markdown",
     "url": "https://github.com/PanQiWei/AutoGPTQ",
-    "keywords": ["gptq", "quantization", "large-language-models", "transformers"],
+    "keywords": ["gptq", "quantization", "large-language-models", "pytorch", "transformers"],
     "platforms": ["windows", "linux"],
     "classifiers": [
         "Environment :: GPU :: NVIDIA CUDA :: 11.7",
         "Environment :: GPU :: NVIDIA CUDA :: 11.8",
-        "Environment :: GPU :: NVIDIA CUDA :: 12",
         "License :: OSI Approved :: MIT License",
         "Natural Language :: Chinese (Simplified)",
         "Natural Language :: English",
         "Programming Language :: Python :: 3.8",
         "Programming Language :: Python :: 3.9",
         "Programming Language :: Python :: 3.10",
-        "Programming Language :: Python :: 3.11",
         "Programming Language :: C++",
-    ]
+    ],
+    "python_requires": f">={python_min_version_str}"
 }
 
-
-PYPI_RELEASE = os.environ.get('PYPI_RELEASE', None)
-BUILD_CUDA_EXT = int(os.environ.get('BUILD_CUDA_EXT', '1')) == 1
 if BUILD_CUDA_EXT:
-    try:
-        import torch
-    except:
-        print("Building cuda extension requires PyTorch(>=1.13.0) been installed, please install PyTorch first!")
-        sys.exit(-1)
-
-    CUDA_VERSION = None
-    ROCM_VERSION = os.environ.get('ROCM_VERSION', None)
-    if ROCM_VERSION and not torch.version.hip:
-        print(
-            f"Trying to compile auto-gptq for RoCm, but PyTorch {torch.__version__} "
-            "is installed without RoCm support."
-        )
-        sys.exit(-1)
-
-    if not ROCM_VERSION:
-        default_cuda_version = torch.version.cuda
-        CUDA_VERSION = "".join(os.environ.get("CUDA_VERSION", default_cuda_version).split("."))
-
     if ROCM_VERSION:
         common_setup_kwargs['version'] += f"+rocm{ROCM_VERSION}"
     else:
-        if not CUDA_VERSION:
-            print(
-                f"Trying to compile auto-gptq for CUDA, byt Pytorch {torch.__version__} "
-                "is installed without CUDA support."
-            )
-            sys.exit(-1)
+        assert CUDA_VERSION
+        common_setup_kwargs['version'] += f"+cu{CUDA_VERSION}"
 
-        # For the PyPI release, the version is simply x.x.x to comply with PEP 440.
-        if not PYPI_RELEASE:
-            common_setup_kwargs['version'] += f"+cu{CUDA_VERSION}"
 
 requirements = [
     "accelerate>=0.19.0",
     "datasets",
-    "sentencepiece",
     "numpy",
     "rouge",
-    "gekko",
     "torch>=1.13.0",
     "safetensors",
     "transformers>=4.31.0",
-    "peft",
-    "tqdm",
+    "peft"
 ]
 
 extras_require = {
-    "triton": ["triton==2.0.0"],
-    "test": ["parameterized"]
+    "triton": ["triton>=2.0.0"]
 }
 
 include_dirs = ["autogptq_cuda"]
@@ -95,14 +80,7 @@ include_dirs = ["autogptq_cuda"]
 additional_setup_kwargs = dict()
 if BUILD_CUDA_EXT:
     from torch.utils import cpp_extension
-       
-    if platform.system() != 'Windows':
-        p = int(subprocess.run("cat /proc/cpuinfo | grep cores | head -1", shell=True, check=True, text=True, stdout=subprocess.PIPE).stdout.split(" ")[2])
-        ret = subprocess.call(["python", "./autogptq_extension/qigen/generate.py", "--module", "--search", "--p", str(p)])
-#        if ret != 0:
-#            raise Exception(f"Failed generate with {ret}")
-#            sys.exit(-1)
-        
+
     if not ROCM_VERSION:
         from distutils.sysconfig import get_python_lib
         conda_cuda_include_dir = os.path.join(get_python_lib(), "nvidia/cuda_runtime/include")
@@ -115,63 +93,28 @@ if BUILD_CUDA_EXT:
         cpp_extension.CUDAExtension(
             "autogptq_cuda_64",
             [
-                "autogptq_extension/cuda_64/autogptq_cuda_64.cpp",
-                "autogptq_extension/cuda_64/autogptq_cuda_kernel_64.cu"
+                "autogptq_cuda/autogptq_cuda_64.cpp",
+                "autogptq_cuda/autogptq_cuda_kernel_64.cu"
             ]
         ),
         cpp_extension.CUDAExtension(
             "autogptq_cuda_256",
             [
-                "autogptq_extension/cuda_256/autogptq_cuda_256.cpp",
-                "autogptq_extension/cuda_256/autogptq_cuda_kernel_256.cu"
+                "autogptq_cuda/autogptq_cuda_256.cpp",
+                "autogptq_cuda/autogptq_cuda_kernel_256.cu"
             ]
-        )
-    ]
-    
-    if platform.system() != 'Windows':
-        extensions.append(
-            cpp_extension.CppExtension(
-                "cQIGen",
-                [
-                    'autogptq_extension/qigen/backend.cpp'
-                ],
-                extra_compile_args = ["-O3", "-mavx", "-mavx2", "-mfma", "-march=native", "-ffast-math", "-ftree-vectorize", "-faligned-new", "-std=c++17", "-fopenmp", "-fno-signaling-nans", "-fno-trapping-math"]
-            )
-        )
-        
-    if os.name == "nt":
-        # On Windows, fix an error LNK2001: unresolved external symbol cublasHgemm bug in the compilation
-        cuda_path = os.environ.get("CUDA_PATH", None)
-        if cuda_path is None:
-            raise ValueError("The environment variable CUDA_PATH must be set to the path to the CUDA install when installing from source on Windows systems.")
-        extra_link_args = ["-L", f"{cuda_path}/lib/x64/cublas.lib"]
-    else:
-        extra_link_args = []
-
-    extensions.append(
+        ),
         cpp_extension.CUDAExtension(
             "exllama_kernels",
             [
-                "autogptq_extension/exllama/exllama_ext.cpp",
-                "autogptq_extension/exllama/cuda_buffers.cu",
-                "autogptq_extension/exllama/cuda_func/column_remap.cu",
-                "autogptq_extension/exllama/cuda_func/q4_matmul.cu",
-                "autogptq_extension/exllama/cuda_func/q4_matrix.cu"
-            ],
-            extra_link_args=extra_link_args
+                "autogptq_cuda/exllama/exllama_ext.cpp",
+                "autogptq_cuda/exllama/cuda_buffers.cu",
+                "autogptq_cuda/exllama/cuda_func/column_remap.cu",
+                "autogptq_cuda/exllama/cuda_func/q4_matmul.cu",
+                "autogptq_cuda/exllama/cuda_func/q4_matrix.cu"
+            ]
         )
-    )
-    extensions.append(
-        cpp_extension.CUDAExtension(
-            "exllamav2_kernels",
-            [
-                "autogptq_extension/exllamav2/ext.cpp",
-                "autogptq_extension/exllamav2/cuda/q_matrix.cu",
-                "autogptq_extension/exllamav2/cuda/q_gemm.cu",
-            ],
-            extra_link_args=extra_link_args
-        )
-    )
+    ]
 
     additional_setup_kwargs = {
         "ext_modules": extensions,
@@ -183,6 +126,5 @@ setup(
     install_requires=requirements,
     extras_require=extras_require,
     include_dirs=include_dirs,
-    python_requires=">=3.8.0",
     **common_setup_kwargs
 )

tests/test_q4.py

@@ -6,9 +6,8 @@ from auto_gptq.utils.import_utils import dynamically_import_QuantLinear
 from auto_gptq.nn_modules.qlinear.qlinear_exllama import QuantLinear
 
 from exllama_kernels import prepare_buffers, set_tuning_params
-from auto_gptq import AutoGPTQForCausalLM, exllama_set_max_input_length
+from auto_gptq import AutoGPTQForCausalLM
 from auto_gptq.modeling._utils import autogptq_post_init
-from auto_gptq.modeling._const import EXLLAMA_DEFAULT_MAX_INPUT_LENGTH
 
 from transformers import AutoTokenizer
 
@@ -143,7 +142,7 @@ class TestsQ4Exllama(unittest.TestCase):
         n = 1024
         device = torch.device("cuda:0")
 
-        linear_class = dynamically_import_QuantLinear(use_triton=False, desc_act=False, group_size=group_size, bits=4, disable_exllama=False, disable_exllamav2=True)
+        linear_class = dynamically_import_QuantLinear(use_triton=False, desc_act=False, group_size=group_size, bits=4)
 
         linear = linear_class(
             bits=4,
@@ -189,35 +188,6 @@ class TestsQ4Exllama(unittest.TestCase):
 
         self.assertTrue(torch.allclose(res, reference, rtol=3e-5, atol=2e-2), get_diff(res, reference))
     
-    def test_exllama_buffer_size(self):
-        prompt = "I am in Paris and" * 450
-        device = torch.device("cuda:0")
-
-        model_id = "TheBloke/vicuna-13B-1.1-GPTQ-4bit-128g"
-        revision = "actorder"
-        model_basename = "vicuna-13B-1.1-GPTQ-4bit-128g.latest"
-
-        model_q = AutoGPTQForCausalLM.from_quantized(model_id, revision=revision, device="cuda:0", use_triton=False, use_safetensors=True, inject_fused_attention=False, inject_fused_mlp=True, model_basename=model_basename, disable_exllama=False, disable_exllamav2=True)
-        tokenizer = AutoTokenizer.from_pretrained(model_id)
-
-        inp = tokenizer(prompt, return_tensors="pt").to(device)
-
-        self.assertTrue(inp["input_ids"].shape[1] > EXLLAMA_DEFAULT_MAX_INPUT_LENGTH)  # 2048 is the default max_input_length
-
-        with self.assertRaises(RuntimeError) as cm:
-            res = model_q.generate(**inp, num_beams=1, min_new_tokens=3, max_new_tokens=3)
-        self.assertTrue("temp_state buffer is too small" in str(cm.exception))
-
-        model_q = exllama_set_max_input_length(model_q, 4096)
-
-        res = model_q.generate(**inp, num_beams=1, min_new_tokens=3, max_new_tokens=3)
-
-        model_q = exllama_set_max_input_length(model_q, 1034)
-
-        with self.assertRaises(RuntimeError) as cm:
-            res = model_q.generate(**inp, num_beams=1, min_new_tokens=3, max_new_tokens=3)
-        self.assertTrue("temp_state buffer is too small" in str(cm.exception))
-
     def test_generation_no_act_order(self):
         prompt = "I am in Paris and"
         device = torch.device("cuda:0")
@@ -226,8 +196,8 @@ class TestsQ4Exllama(unittest.TestCase):
         reference_output = "<s> I am in Paris and I am going to the Louvre Museum. What time does it open and what is the best way to get there?\nThe Louvre Museum in Paris is open from 9:00 AM to 6:00 PM every day except for Tuesdays. The best way to get"
 
         model_id = "TheBloke/WizardLM-7B-uncensored-GPTQ"
-        model_basename = "model"
-        model_q = AutoGPTQForCausalLM.from_quantized(model_id, device="cuda:0", use_triton=False, use_safetensors=True, inject_fused_attention=True, inject_fused_mlp=True, model_basename=model_basename, disable_exllama=False, disable_exllamav2=True)
+        model_basename = "WizardLM-7B-uncensored-GPTQ-4bit-128g.compat.no-act-order"
+        model_q = AutoGPTQForCausalLM.from_quantized(model_id, device="cuda:0", use_triton=False, use_safetensors=True, inject_fused_attention=True, inject_fused_mlp=True, model_basename=model_basename)
         tokenizer = AutoTokenizer.from_pretrained(model_id)
 
         inp = tokenizer(prompt, return_tensors="pt").to(device)
@@ -249,7 +219,7 @@ class TestsQ4Exllama(unittest.TestCase):
         revision = "actorder"
         model_basename = "vicuna-13B-1.1-GPTQ-4bit-128g.latest"
 
-        model_q = AutoGPTQForCausalLM.from_quantized(model_id, revision=revision, device="cuda:0", use_triton=False, use_safetensors=True, inject_fused_attention=False, inject_fused_mlp=True, model_basename=model_basename, disable_exllama=False, disable_exllamav2=True)
+        model_q = AutoGPTQForCausalLM.from_quantized(model_id, revision=revision, device="cuda:0", use_triton=False, use_safetensors=True, inject_fused_attention=False, inject_fused_mlp=True, model_basename=model_basename, disable_exllama=False)
         tokenizer = AutoTokenizer.from_pretrained(model_id)
 
         inp = tokenizer(prompt, return_tensors="pt").to(device)
@@ -338,7 +308,7 @@ class TestsQ4CUDA(unittest.TestCase):
         n = 256
         device = "cuda"
 
-        linear_class = dynamically_import_QuantLinear(use_triton=False, desc_act=False, group_size=group_size, bits=4, disable_exllamav2=True)
+        linear_class = dynamically_import_QuantLinear(use_triton=False, desc_act=False, group_size=group_size, bits=4, disable_exllama=True)
 
         linear = linear_class(
             bits=4,
@@ -369,226 +339,3 @@ class TestsQ4CUDA(unittest.TestCase):
             reference = self.REFERENCE_OLD_NO_HALF.to(device)
 
         self.assertTrue(torch.allclose(res, reference), get_diff(res, reference))
-
-
-class TestsQ4ExllamaV2(unittest.TestCase):
-
-    # reference generated with cuda_old
-    REFERENCE = torch.Tensor([5.8398, 6.8555, 7.2734, 6.4219, 6.2070, 5.8203, 6.5664, 6.4219, 6.2148,
-        5.3281, 5.7578, 7.5312, 8.1016, 6.1133, 7.2031, 6.6484, 6.5156, 6.0117,
-        6.0312, 6.1914, 6.2109, 6.8125, 5.8125, 7.1172, 7.3125, 6.7305, 5.9961,
-        6.5117, 6.1914, 5.9648, 7.1680, 6.4766, 7.2070, 6.5469, 6.7734, 6.4219,
-        6.8086, 7.0469, 5.9297, 6.4727, 6.2539, 5.9570, 7.2383, 5.8945, 6.0820,
-        5.7969, 7.1094, 6.2188, 6.7500, 7.3555, 6.2930, 6.7734, 5.9219, 7.4805,
-        6.8750, 6.4102, 6.5898, 6.5469, 7.6016, 6.7461, 5.9492, 7.2227, 5.8164,
-        5.4570, 6.2930, 7.3984, 6.0938, 7.3984, 5.9609, 6.3516, 6.5664, 5.7969,
-        7.1250, 6.0781, 6.7930, 5.9492, 6.1641, 6.5898, 6.0586, 6.3359, 6.7930,
-        7.0469, 6.0664, 6.3320, 5.4414, 6.7617, 5.1641, 7.2891, 6.8516, 6.5312,
-        5.6914, 7.3711, 6.8203, 5.9492, 7.0781, 6.3164, 7.1992, 7.1133, 7.4219,
-        7.5586, 7.1836, 6.9102, 6.4844, 6.9805, 6.1953, 6.5156, 5.4844, 6.6602,
-        6.6719, 7.9844, 6.4727, 6.6367, 6.2227, 6.4531, 5.0625, 6.4609, 6.7031,
-        6.6445, 6.5234, 6.8633, 6.6055, 5.6055, 6.4453, 7.2617, 6.3945, 6.6367,
-        6.1055, 7.0664, 6.0820, 6.6875, 6.1445, 6.8672, 6.2070, 6.8828, 6.1484,
-        6.7070, 6.8516, 6.2734, 7.1055, 7.0586, 6.9648, 5.9727, 6.1016, 6.8750,
-        7.0078, 7.1523, 5.7383, 5.9531, 6.5508, 7.5352, 6.1602, 6.2578, 6.3906,
-        5.7383, 6.7031, 5.7344, 6.3516, 5.2852, 7.5312, 6.4531, 6.6406, 6.2266,
-        6.1094, 5.9102, 5.7617, 6.3789, 7.0508, 6.3750, 6.3320, 6.8555, 6.7266,
-        7.0352, 7.7695, 6.3984, 6.5039, 6.8320, 6.1602, 6.0312, 6.3828, 6.9023,
-        7.4336, 7.3711, 6.1016, 7.0703, 6.3281, 6.8281, 6.4922, 5.9453, 5.1016,
-        6.7188, 6.1406, 6.6289, 7.2695, 6.2070, 6.7070, 7.2930, 7.1836, 6.3828,
-        6.1992, 6.7070, 7.8008, 7.7773, 5.6602, 7.0273, 6.6172, 6.0898, 5.3516,
-        7.3359, 5.9727, 6.0078, 7.0586, 6.3086, 6.8555, 7.2617, 7.3477, 6.3828,
-        7.1133, 6.6328, 7.3516, 6.9141, 7.2031, 6.9805, 6.1719, 6.7812, 8.3047,
-        6.5898, 6.3633, 6.2539, 7.2773, 6.5938, 6.4141, 6.8203, 6.8906, 7.8828,
-        5.9609, 6.4180, 7.3984, 5.7539, 7.1758, 6.6641, 6.9062, 6.2578, 7.5508,
-        6.1719, 6.5742, 5.9375, 6.7891, 6.2109, 6.5039, 6.8750, 6.2031, 6.8828,
-        7.1094, 5.9570, 7.2969, 6.6797, 6.8828, 5.5430, 6.9648, 5.8398, 6.5430,
-        6.3945, 6.5664, 5.8086, 6.6172, 7.0586, 6.8867, 6.0820, 5.8125, 6.7070,
-        7.5742, 6.2578, 6.1328, 6.5391, 5.4531, 6.8242, 6.6953, 6.8008, 6.3398,
-        6.4805, 7.2266, 6.3281, 6.6875, 6.4688, 5.9414, 7.4297, 5.8711, 6.0625,
-        5.8750, 6.5664, 5.8867, 6.3477, 6.1133, 6.9453, 5.0547, 6.7812, 6.4922,
-        7.2422, 5.4688, 6.2109, 7.2148, 6.1758, 5.9297, 7.1953, 5.5195, 6.3203,
-        5.9961, 7.9297, 6.2695, 6.4414, 6.7266, 7.1875, 7.3203, 5.4062, 6.0625,
-        7.0898, 5.3828, 5.6133, 6.0742, 6.6836, 5.7109, 7.2852, 7.7539, 7.5820,
-        6.4258, 5.9336, 6.3750, 6.3555, 7.5469, 6.2539, 6.5898, 6.4102, 7.0469,
-        5.7344, 7.2031, 6.7969, 5.6836, 7.6523, 6.9297, 7.8672, 6.4766, 6.3008,
-        7.0977, 6.5430, 7.0938, 5.8398, 6.9883, 6.5312, 6.3203, 6.3594, 5.4062,
-        6.9688, 5.7930, 6.3164, 6.5547, 7.1992, 5.8750, 6.3008, 6.7930, 6.0391,
-        7.4766, 6.6094, 6.5625, 5.9805, 6.2422, 7.2109, 6.6875, 5.3047, 7.6211,
-        5.9453, 6.5625, 6.1641, 6.1250, 6.5977, 7.7422, 7.0742, 5.6875, 6.2656,
-        6.6250, 6.8945, 5.7070, 6.3203, 5.7500, 6.2695, 6.2773, 6.8516, 6.4883,
-        7.0000, 6.7578, 6.1875, 5.9844, 5.5703, 6.7188, 5.5273, 5.3438, 7.2500,
-        6.7852, 6.5195, 6.8125, 6.0664, 6.7852, 7.0000, 7.0781, 6.8477, 7.2930,
-        6.3438, 7.1523, 6.3281, 6.8047, 7.3203, 5.3359, 6.1484, 6.5586, 7.3828,
-        6.2344, 7.1523, 6.4102, 5.5898, 7.0195, 7.1172, 5.8008, 6.5742, 6.2891,
-        8.0312, 6.9023, 6.5898, 7.1953, 6.7266, 6.0078, 5.5430, 6.4766, 6.4258,
-        5.9648, 8.0859, 5.0547, 7.2188, 7.4375, 6.5156, 5.9922, 6.3281, 6.2852,
-        6.7734, 6.2461, 6.9805, 5.4648, 5.8867, 6.8242, 6.3008, 6.3281, 7.3047,
-        7.1836, 6.5195, 6.6328, 6.7188, 5.4336, 6.5078, 5.3477, 5.5508, 7.3125,
-        5.8750, 6.5195, 6.2383, 6.3594, 6.0898, 6.4141, 5.9844, 6.6250, 7.7109,
-        6.0391, 7.2344, 5.9453, 5.9453, 7.0586, 5.6641, 7.2773, 6.5195, 7.2227,
-        6.3359, 5.3203, 6.4375, 7.2383, 6.4023, 6.2148, 7.3750, 5.8164, 6.2109,
-        6.5430, 5.8164, 6.1680, 6.7656, 6.0820, 6.1094, 6.5312, 6.8906, 6.8320,
-        6.1289, 6.3125, 7.6797, 6.3008, 6.0000, 7.3320, 6.7852, 6.9297, 6.6328,
-        6.2266, 5.1602, 6.2031, 7.0547, 5.9492, 6.0703, 6.0977, 6.8086, 6.0742,
-        6.0195, 7.0625, 6.5781, 5.7461, 6.1562, 7.0430, 6.7148, 6.5312, 6.5820,
-        6.4570, 7.5508, 5.6289, 6.0547, 6.5000, 7.3125, 5.8477, 5.9297, 6.2578,
-        6.0078, 5.9922, 7.3398, 7.4922, 7.8906, 7.5547, 5.4648, 6.5156, 6.3242,
-        6.1094, 6.9219, 6.7227, 6.6836, 7.4023, 5.9648, 7.2383, 6.7695, 6.6797,
-        7.0547, 6.3047, 6.4688, 6.9961, 6.0391, 5.9727, 6.8398, 6.7422, 5.7656,
-        5.4766, 6.7852, 7.0820, 5.3516, 7.6523, 5.1562, 6.6445, 6.1211, 6.2695,
-        6.0703, 6.3594, 6.4062, 6.3398, 5.7578, 6.5391, 6.2500, 6.5742, 6.5000,
-        7.5625, 7.0117, 6.5547, 7.1250, 6.4453, 6.6094, 6.1875, 6.4219, 6.6172,
-        6.4336, 6.5703, 6.1758, 6.4219, 6.6016, 6.7383, 6.7070, 6.1328, 5.5586,
-        6.6367, 6.3789, 6.2578, 5.5039, 6.6172, 6.4648, 5.8086, 7.2031, 5.8125,
-        6.3711, 7.6758, 7.1289, 5.8086, 6.3008, 6.2109, 6.1602, 6.1797, 7.2305,
-        6.7266, 6.2422, 5.6719, 6.7070, 6.9414, 6.8594, 7.4023, 7.2109, 6.0156,
-        6.6680, 6.6172, 7.1250, 6.6523, 6.9531, 6.7617, 6.4961, 6.9414, 5.7188,
-        7.6367, 6.5469, 6.2305, 6.4414, 7.4648, 5.9102, 6.2461, 6.1367, 6.8203,
-        6.5703, 6.8867, 7.0000, 6.7539, 6.1719, 6.5469, 6.2422, 5.4297, 5.7305,
-        5.1641, 6.1875, 7.0312, 6.6484, 6.0234, 7.4102, 6.8711, 6.3086, 6.3711,
-        6.7344, 6.6992, 5.9766, 7.3906, 7.1875, 6.4883, 6.3984, 7.3438, 6.9688,
-        6.9062, 6.4375, 6.7891, 7.0117, 6.4883, 5.7500, 7.0898, 7.0742, 6.7070,
-        5.8750, 6.0469, 6.6445, 5.2773, 6.8984, 6.1641, 7.0508, 7.4609, 5.0273,
-        6.7734, 6.4531, 5.7656, 6.5312, 7.4648, 6.1250, 6.5625, 7.1367, 6.0625,
-        6.1211, 6.9766, 6.6758, 6.3164, 6.8828, 6.8203, 6.7500, 6.5352, 7.3008,
-        6.7852, 6.1914, 5.0508, 6.7188, 7.1172, 6.8008, 6.8086, 5.4883, 6.9180,
-        6.5742, 6.1719, 7.0469, 7.1523, 5.9492, 5.8594, 6.8320, 6.1719, 6.2031,
-        6.8398, 7.3008, 6.6289, 6.4922, 6.0000, 5.4766, 6.3320, 6.5117, 6.2812,
-        7.5742, 6.3516, 7.0039, 6.4570, 7.1523, 7.6289, 6.2578, 7.1875, 6.4844,
-        5.7930, 6.7070, 7.5508, 7.1797, 6.0430, 6.8711, 6.5742, 7.5781, 6.4766,
-        6.5391, 6.9453, 6.1992, 6.6367, 6.2812, 6.0234, 6.6953, 7.0312, 6.2031,
-        6.5625, 6.6719, 6.1719, 6.5586, 5.7031, 7.4609, 6.6211, 7.7227, 6.9141,
-        6.0469, 6.2500, 5.3828, 6.0078, 5.8164, 5.8867, 6.1523, 6.6523, 6.6953,
-        7.3125, 6.4844, 5.9570, 5.9531, 6.2109, 5.5039, 6.5117, 6.8203, 6.6133,
-        6.4766, 5.9297, 7.1445, 7.1914, 6.0117, 6.8281, 6.7422, 6.1328, 6.9805,
-        6.5625, 6.9180, 7.1133, 7.3359, 5.7617, 5.8711, 6.4961, 6.5859, 6.2422,
-        6.5273, 6.7461, 6.6992, 6.7695, 6.6289, 5.9453, 5.9805, 7.1172, 6.6719,
-        6.0039, 7.6875, 6.7812, 7.8359, 6.9531, 7.4336, 7.6602, 6.8164, 7.3945,
-        7.1602, 6.8789, 5.0078, 6.0547, 6.8086, 6.7070, 6.4688, 6.4492, 6.6172,
-        5.5625, 6.6914, 6.4297, 5.7461, 5.3359, 6.8750, 6.4609, 7.4062, 5.2070,
-        6.0820, 6.7383, 6.5703, 6.1797, 6.7070, 6.5977, 5.9961, 6.6328, 6.9375,
-        6.3906, 6.6484, 4.9609, 6.6445, 6.5898, 7.1875, 7.5195, 6.7969, 6.1367,
-        6.8906, 7.4297, 6.3633, 6.0508, 6.5000, 6.4648, 6.7539, 6.7109, 5.8086,
-        6.6016, 7.1133, 4.8672, 6.6367, 6.1641, 5.1758, 6.9453, 6.3242, 7.0664,
-        6.4805, 6.3516, 6.7383, 8.4688, 6.7305, 5.9844, 6.5938, 7.2969, 6.5977,
-        7.5898, 6.2969, 6.8672, 6.6680, 7.1289, 6.6875, 5.4258, 8.1875, 8.0391,
-        7.7969, 6.6445, 7.0703, 7.3359, 6.9805, 6.6328, 6.5352, 6.2422, 5.5820,
-        6.8633, 6.8047, 6.5703, 6.0117, 6.7539, 7.1719, 6.8438, 7.3633, 6.6016,
-        7.2070, 6.4727, 5.8008, 7.4062, 7.4805, 6.6445, 5.9023, 6.3984, 6.9961,
-        6.6680, 6.8242, 6.7148, 6.6172, 6.9727, 6.8320, 5.9766, 6.6133, 5.5977,
-        6.7773, 7.3906, 6.9219, 7.0781, 6.6914, 5.7539, 6.7969, 6.8008, 5.8047,
-        7.1055, 6.4961, 6.0352, 5.6211, 7.4414, 7.0703, 6.1172, 6.7461, 6.4492,
-        7.7148, 6.4258, 6.0039, 6.5156, 7.2188, 7.4531, 7.4844, 7.5938, 7.4023,
-        6.7617, 6.0078, 6.3320, 5.8906, 7.5977, 5.6523, 6.7734, 6.3008, 5.2227,
-        7.1719, 7.1289, 6.6602, 5.4609, 7.0312, 6.0820, 6.1719, 6.0000, 6.5547,
-        6.6328, 7.0547, 7.0859, 6.2656, 5.5234, 6.0273, 6.7891, 7.1875, 6.9531,
-        6.8203, 6.3516, 6.1172, 6.4648, 6.9180, 7.3906, 6.2812, 5.7109, 6.1484,
-        6.9102, 6.8711, 7.0156, 6.1445, 5.8867, 6.3828, 5.9961, 6.6914, 6.7891,
-        7.0820, 6.6719, 6.9297, 6.3750, 6.7578, 6.4883, 6.2227, 6.2305, 6.0508,
-        6.6484, 5.7578, 7.2070, 7.2383, 6.9375, 7.2578, 6.5312, 6.0312, 6.7930,
-        6.2578, 7.0625, 7.2148, 6.4961, 7.0703, 6.4727, 7.3906]).to(torch.float16)
-
-    def test_exllamav2(self):
-        from auto_gptq.nn_modules.qlinear.qlinear_exllamav2 import QuantLinear
-        
-        group_size = 128
-
-        m = 1
-        k = 1024
-        n = 1024
-        device = torch.device("cuda:0")
-
-        linear_class = dynamically_import_QuantLinear(use_triton=False, desc_act=False, group_size=group_size, bits=4)
-
-        linear = linear_class(
-            bits=4,
-            group_size=group_size,
-            infeatures=k,
-            outfeatures=n,
-            bias=False,
-        )
-        
-        self.assertTrue(isinstance(linear, QuantLinear))
-
-        torch.manual_seed(42)
-
-        linear.qweight = torch.randint(-100, 100, size=linear.qweight.shape, dtype=torch.int32)
-        linear.scales = linear.scales + 0.002
-
-        linear = linear.eval()
-        linear = linear.to(device)
-
-        linear = autogptq_post_init(linear, use_act_order=False)
-
-        inp = torch.rand(1, m, k, dtype=torch.float16).to(device)
-
-        with torch.no_grad():
-            res = linear(inp)[0][0]
-
-        reference = self.REFERENCE.to(device)
-
-        self.assertTrue(torch.allclose(res, reference, rtol=3e-5, atol=2e-2), get_diff(res, reference))
-    
-    def test_generation_no_act_order(self):
-        prompt = "I am in Paris and"
-        device = torch.device("cuda:0")
-
-        # Reference generated with the cuda-old kernel
-        reference_output = "<s> I am in Paris and I am going to the Louvre Museum. What time does it open and what is the best way to get there?\nThe Louvre Museum in Paris is open from 9:00 AM to 6:00 PM every day except for Tuesdays. The best way to get"
-        
-        model_id = "TheBloke/WizardLM-7B-uncensored-GPTQ"
-        model_basename = "model"
-        
-        model_q = AutoGPTQForCausalLM.from_quantized(model_id, device="cuda:0", use_triton=False, use_safetensors=True, model_basename=model_basename)
-        tokenizer = AutoTokenizer.from_pretrained(model_id)
-
-        inp = tokenizer(prompt, return_tensors="pt").to(device)
-
-        res = model_q.generate(**inp, num_beams=1, min_new_tokens=60, max_new_tokens=60)
-
-        predicted_text = tokenizer.decode(res[0])
-        
-
-        self.assertEqual(predicted_text, reference_output)
-        
-    def test_generation_with_act_order(self):
-        prompt = "I am in Paris and"
-        device = torch.device("cuda:0")
-
-        # Reference generated with the cuda-old kernel
-        reference_output = "<s> I am in Paris and it is a beautiful day. I am sitting in a café, drinking coffee and writing this book. I am surrounded by the sights and sounds of the city, and I am filled with a sense of contentment and gratitude.\n\nI am grateful for the opportunity to live and"
-
-        model_id = "TheBloke/vicuna-13B-1.1-GPTQ-4bit-128g"
-        revision = "actorder"
-        model_basename = "vicuna-13B-1.1-GPTQ-4bit-128g.latest"
-
-        model_q = AutoGPTQForCausalLM.from_quantized(model_id, revision=revision, device="cuda:0", use_triton=False, use_safetensors=True, inject_fused_attention=False, inject_fused_mlp=True, model_basename=model_basename)
-        tokenizer = AutoTokenizer.from_pretrained(model_id)
-
-        inp = tokenizer(prompt, return_tensors="pt").to(device)
-
-        res = model_q.generate(**inp, num_beams=1, min_new_tokens=60, max_new_tokens=60)
-
-        predicted_text = tokenizer.decode(res[0])
-
-        self.assertEqual(predicted_text, reference_output)
-
-    def test_exllama_buffer_size(self):
-        # prompt = "I'm in Paris and" * 450
-        prompt = "I'm in Paris and" * 1000
-        device = torch.device("cuda:0")
-
-        model_id = "TheBloke/vicuna-13B-1.1-GPTQ-4bit-128g"
-        revision = "actorder"
-        model_basename = "vicuna-13B-1.1-GPTQ-4bit-128g.latest"
-
-        model_q = AutoGPTQForCausalLM.from_quantized(model_id, revision=revision, device="cuda:0", use_triton=False, use_safetensors=True, inject_fused_attention=True, inject_fused_mlp=True, model_basename=model_basename)
-        
-        tokenizer = AutoTokenizer.from_pretrained(model_id)
-
-        inp = tokenizer(prompt, return_tensors="pt").to(device)
-        
-        self.assertTrue(inp["input_ids"].shape[1] > 2048)  # 2048 is the default max_input_length for LLama
-        
-        res = model_q.generate(**inp, num_beams=1, min_new_tokens=3, max_new_tokens=3)